Duplicate from editing-images/ledits

Co-authored-by: Linoy Tsaban <[email protected]>

.gitattributes

@@ -0,0 +1,37 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+exmaples/source_a_man_wearing_a_brown_hoodie_in_a_crowded_street.jpeg filter=lfs diff=lfs merge=lfs -text
+examples/source_a_man_wearing_a_brown_hoodie_in_a_crowded_street.jpeg filter=lfs diff=lfs merge=lfs -text
+examples/source_wall_with_framed_photos.jpeg filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,20 @@
+---
+title: LEDITS
+emoji: ✏️
+colorFrom: gray
+colorTo: blue
+sdk: gradio
+sdk_version: 3.36.1
+app_file: app.py
+pinned: true
+duplicated_from: editing-images/ledits
+---
+
+This is the repository for [LEDITS - Real Image Latent Editing with Edit Friendly DDPM and Semantic Guidance](arxiv.org/abs/2307.00522).  More information about the technique [here](https://editing-images-project.hf.space)
+
+This repository contains the following relevant files: 
+- `app.py` - Gradio application for the inversion technique combining uploading an image, captioning it, doing the DDPM Inversion and applying SEGA concepts to the editing.
+- `constants.py` - default config values for the `app.py`
+- `inversion_utils.py` - utilities for providing the DDPM Inversion
+- `modified_pipeline_semantic_stable_diffusion.py` - modified pipeline of SEGA for the purposes of LEDITS
+- `utils.py` - generic useful utils for the app

app.py

@@ -0,0 +1,867 @@
+import gradio as gr
+import torch
+import numpy as np
+import requests
+import random
+from io import BytesIO
+from utils import *
+from constants import *
+from inversion_utils import *
+from modified_pipeline_semantic_stable_diffusion import SemanticStableDiffusionPipeline
+from torch import autocast, inference_mode
+from diffusers import StableDiffusionPipeline
+from diffusers import DDIMScheduler
+from transformers import AutoProcessor, BlipForConditionalGeneration
+from share_btn import community_icon_html, loading_icon_html, share_js
+
+# load pipelines
+sd_model_id = "stabilityai/stable-diffusion-2-1-base"
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+sd_pipe = StableDiffusionPipeline.from_pretrained(sd_model_id,torch_dtype=torch.float16).to(device)
+sd_pipe.scheduler = DDIMScheduler.from_config(sd_model_id, subfolder = "scheduler")
+sem_pipe = SemanticStableDiffusionPipeline.from_pretrained(sd_model_id, torch_dtype=torch.float16).to(device)
+blip_processor = AutoProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
+blip_model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base",torch_dtype=torch.float16).to(device)
+
+
+
+## IMAGE CPATIONING ##
+def caption_image(input_image):
+    inputs = blip_processor(images=input_image, return_tensors="pt").to(device, torch.float16)
+    pixel_values = inputs.pixel_values
+
+    generated_ids = blip_model.generate(pixel_values=pixel_values, max_length=50)
+    generated_caption = blip_processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
+    return generated_caption, generated_caption
+
+
+
+## DDPM INVERSION AND SAMPLING ##
+def invert(x0, prompt_src="", num_diffusion_steps=100, cfg_scale_src = 3.5, eta = 1):
+
+  #  inverts a real image according to Algorihm 1 in https://arxiv.org/pdf/2304.06140.pdf,
+  #  based on the code in https://github.com/inbarhub/DDPM_inversion
+
+  #  returns wt, zs, wts:
+  #  wt - inverted latent
+  #  wts - intermediate inverted latents
+  #  zs - noise maps
+
+  sd_pipe.scheduler.set_timesteps(num_diffusion_steps)
+
+  # vae encode image
+  with inference_mode():
+    w0 = (sd_pipe.vae.encode(x0).latent_dist.mode() * 0.18215)
+
+  # find Zs and wts - forward process
+  wt, zs, wts = inversion_forward_process(sd_pipe, w0, etas=eta, prompt=prompt_src, cfg_scale=cfg_scale_src, prog_bar=True, num_inference_steps=num_diffusion_steps)
+  return zs, wts
+
+
+def sample(zs, wts, prompt_tar="", cfg_scale_tar=15, skip=36, eta = 1):
+
+    # reverse process (via Zs and wT)
+    w0, _ = inversion_reverse_process(sd_pipe, xT=wts[skip], etas=eta, prompts=[prompt_tar], cfg_scales=[cfg_scale_tar], prog_bar=True, zs=zs[skip:])
+
+    # vae decode image
+    with inference_mode():
+      x0_dec = sd_pipe.vae.decode(1 / 0.18215 * w0).sample
+    if x0_dec.dim()<4:
+        x0_dec = x0_dec[None,:,:,:]
+    img = image_grid(x0_dec)
+    return img
+
+
+def reconstruct(tar_prompt,
+                image_caption,
+                tar_cfg_scale,
+                skip,
+                wts, zs,
+                do_reconstruction,
+                reconstruction,
+                reconstruct_button
+               ):
+
+    if reconstruct_button == "Hide Reconstruction":
+      return reconstruction.value, reconstruction, ddpm_edited_image.update(visible=False), do_reconstruction, "Show Reconstruction"
+
+    else:
+      if do_reconstruction:
+          if image_caption.lower() == tar_prompt.lower(): # if image caption was not changed, run actual reconstruction
+             tar_prompt = ""
+          reconstruction_img = sample(zs.value, wts.value, prompt_tar=tar_prompt, skip=skip, cfg_scale_tar=tar_cfg_scale)
+          reconstruction = gr.State(value=reconstruction_img)
+          do_reconstruction = False
+      return reconstruction.value, reconstruction, ddpm_edited_image.update(visible=True), do_reconstruction, "Hide Reconstruction"
+
+
+def load_and_invert(
+                    input_image,
+                    do_inversion,
+                    seed, randomize_seed,
+                    wts, zs,
+                    src_prompt ="",
+                    tar_prompt="",
+                    steps=100,
+                    src_cfg_scale = 3.5,
+                    skip=36,
+                    tar_cfg_scale=15,
+                    progress=gr.Progress(track_tqdm=True)
+
+):
+
+
+    x0 = load_512(input_image, device=device).to(torch.float16)
+
+    if do_inversion or randomize_seed:
+        # invert and retrieve noise maps and latent
+        zs_tensor, wts_tensor = invert(x0 =x0 , prompt_src=src_prompt, num_diffusion_steps=steps, cfg_scale_src=src_cfg_scale)
+        wts = gr.State(value=wts_tensor)
+        zs = gr.State(value=zs_tensor)
+        do_inversion = False
+
+    return wts, zs, do_inversion, inversion_progress.update(visible=False)
+
+## SEGA ##
+
+def edit(input_image,
+            wts, zs,
+            tar_prompt,
+            image_caption,
+            steps,
+            skip,
+            tar_cfg_scale,
+            edit_concept_1,edit_concept_2,edit_concept_3,
+            guidnace_scale_1,guidnace_scale_2,guidnace_scale_3,
+            warmup_1, warmup_2, warmup_3,
+            neg_guidance_1, neg_guidance_2, neg_guidance_3,
+            threshold_1, threshold_2, threshold_3,
+         do_reconstruction,
+         reconstruction,
+         
+         # for inversion in case it needs to be re computed (and avoid delay):
+         do_inversion,
+         seed, 
+         randomize_seed,
+         src_prompt,
+         src_cfg_scale):
+    show_share_button = gr.update(visible=True)
+    if do_inversion or randomize_seed:
+        x0 = load_512(input_image, device=device).to(torch.float16)
+        # invert and retrieve noise maps and latent
+        zs_tensor, wts_tensor = invert(x0 =x0 , prompt_src=src_prompt, num_diffusion_steps=steps, cfg_scale_src=src_cfg_scale)
+        wts = gr.State(value=wts_tensor)
+        zs = gr.State(value=zs_tensor)
+        do_inversion = False    
+    
+    if image_caption.lower() == tar_prompt.lower(): # if image caption was not changed, run pure sega
+          tar_prompt = ""
+        
+    if edit_concept_1 != "" or edit_concept_2 != "" or edit_concept_3 != "":
+      editing_args = dict(
+      editing_prompt = [edit_concept_1,edit_concept_2,edit_concept_3],
+      reverse_editing_direction = [ neg_guidance_1, neg_guidance_2, neg_guidance_3,],
+      edit_warmup_steps=[warmup_1, warmup_2, warmup_3,],
+      edit_guidance_scale=[guidnace_scale_1,guidnace_scale_2,guidnace_scale_3],
+      edit_threshold=[threshold_1, threshold_2, threshold_3],
+      edit_momentum_scale=0.3,
+      edit_mom_beta=0.6,
+      eta=1,)
+
+      latnets = wts.value[skip].expand(1, -1, -1, -1)
+      sega_out = sem_pipe(prompt=tar_prompt, latents=latnets, guidance_scale = tar_cfg_scale,
+                          num_images_per_prompt=1,
+                          num_inference_steps=steps,
+                          use_ddpm=True,  wts=wts.value, zs=zs.value[skip:], **editing_args)
+      
+      return sega_out.images[0], reconstruct_button.update(visible=True), do_reconstruction, reconstruction, wts, zs, do_inversion, show_share_button
+    
+    else: # if sega concepts were not added, performs regular ddpm sampling
+      
+      if do_reconstruction: # if ddpm sampling wasn't computed
+          pure_ddpm_img = sample(zs.value, wts.value, prompt_tar=tar_prompt, skip=skip, cfg_scale_tar=tar_cfg_scale)
+          reconstruction = gr.State(value=pure_ddpm_img)
+          do_reconstruction = False
+          return pure_ddpm_img, reconstruct_button.update(visible=False), do_reconstruction, reconstruction, wts, zs, do_inversion, show_share_button
+      
+      return reconstruction.value, reconstruct_button.update(visible=False), do_reconstruction, reconstruction, wts, zs, do_inversion, show_share_button
+        
+
+def randomize_seed_fn(seed, randomize_seed):
+    if randomize_seed:
+        seed = random.randint(0, np.iinfo(np.int32).max)
+    torch.manual_seed(seed)
+    return seed
+
+def crop_image(image):
+    h, w, c = image.shape
+    if h < w:
+        offset = (w - h) // 2
+        image = image[:, offset:offset + h]
+    elif w < h:
+        offset = (h - w) // 2
+        image = image[offset:offset + w]
+    image = np.array(Image.fromarray(image).resize((512, 512)))
+    return image
+    
+
+def get_example():
+    case = [
+        [
+            'examples/lemons_input.jpg', 
+            # '',
+            'apples', 'lemons',
+            'a ceramic bowl',
+             'examples/lemons_output.jpg',
+            
+            
+            7,7,
+            1,1,
+            False, True,
+            100,
+            36,
+            15,
+           
+             ],
+        [
+            'examples/girl_with_pearl_earring_input.png', 
+            # '',
+            'glasses', '',
+            '',
+             'examples/girl_with_pearl_earring_output.png',
+            
+            
+            3,7,
+            3,2,
+            False,False,
+            100,
+            36,
+            15,
+           
+             ],
+        [
+            'examples/rockey_shore_input.jpg', 
+            # '',
+            'sea turtle', '',
+            'watercolor painting',
+            'examples/rockey_shore_output.jpg',
+            
+            
+            7,7,
+            1,2,
+            False,False,
+            100,
+            36,
+            15,
+             ],
+                 [
+            'examples/flower_field_input.jpg', 
+            # '',
+            'wheat', 'red flowers',
+            'oil painting',
+             'examples/flower_field_output_2.jpg',
+
+
+            20,7,
+            1,1,
+                     False,True,
+                      100,
+            36,
+            15,
+            
+             ],
+        [
+            'examples/butterfly_input.jpg', 
+            # '',
+             'bee', 'butterfly',
+            'oil painting',
+            'examples/butterfly_output.jpg',      
+            7, 7,
+            1,1,
+            False, True,
+                        100,
+            36,
+            15,
+             ]
+ ]
+    return case
+
+
+def swap_visibilities(input_image,  
+                    edit_concept_1,
+                    edit_concept_2,
+                    tar_prompt,  
+                    sega_edited_image,
+                    guidnace_scale_1,
+                    guidnace_scale_2,
+                    warmup_1,
+                    warmup_2,
+                    neg_guidance_1,
+                    neg_guidance_2,
+                    steps,
+                    skip,
+                    tar_cfg_scale,
+                    sega_concepts_counter
+                    
+):
+    sega_concepts_counter=0
+    concept1_update = update_display_concept("Remove" if neg_guidance_1 else "Add", edit_concept_1, neg_guidance_1, sega_concepts_counter)
+    if(edit_concept_2 != ""):
+        concept2_update = update_display_concept("Remove" if neg_guidance_2 else "Add", edit_concept_2, neg_guidance_2, sega_concepts_counter+1)
+    else:
+        concept2_update = gr.update(visible=False), gr.update(visible=False),gr.update(visible=False), gr.update(value=neg_guidance_2),gr.update(visible=True),gr.update(visible=False),sega_concepts_counter+1
+    
+    return (gr.update(visible=True), *concept1_update[:-1], *concept2_update)
+    
+
+
+########
+# demo #
+########
+
+
+intro = """
+<h1 style="font-weight: 1400; text-align: center; margin-bottom: 7px;">
+   LEDITS - Pipeline for editing images
+</h1>
+<h3 style="font-weight: 600; text-align: center;">
+    Real Image Latent Editing with Edit Friendly DDPM and Semantic Guidance
+</h3>
+<h4 style="text-align: center; margin-bottom: 7px;">
+    <a href="https://editing-images-project.hf.space/" style="text-decoration: underline;" target="_blank">Project Page</a> | <a href="https://arxiv.org/abs/2307.00522" style="text-decoration: underline;" target="_blank">ArXiv</a>
+</h4>
+
+<p style="font-size: 0.9rem; margin: 0rem; line-height: 1.2em; margin-top:1em">
+<a href="https://huggingface.co/spaces/editing-images/edit_friendly_ddpm_x_sega?duplicate=true">
+<img style="margin-top: 0em; margin-bottom: 0em" src="https://bit.ly/3CWLGkA" alt="Duplicate Space"></a>
+<p/>"""
+
+help_text = """
+- **Getting Started - edit images with DDPM X SEGA:**
+
+    The are 3 general setting options you can play with -
+
+    1. **Pure DDPM Edit -** Describe the desired edited output image in detail
+    2. **Pure SEGA Edit -** Keep the target prompt empty ***or*** with a description of the original image and add editing concepts for Semantic Gudiance editing
+    3. **Combined -** Describe the desired edited output image in detail and add additional SEGA editing concepts on top
+- **Getting Started - Tips**
+
+    While the best approach depends on your editing objective and source image,  we can layout a few guiding tips to use as a starting point -
+
+    1. **DDPM** is usually more suited for scene/style changes and major subject changes (for example ) while **SEGA** allows for more fine grained control, changes are more delicate, more suited for adding details (for example facial expressions and attributes, subtle style modifications, object adding/removing)
+    2. The more you describe the scene in the target prompt (both the parts and details you wish to keep the same and those you wish to change), the better the result
+    3. **Combining DDPM Edit with SEGA -**
+    Try dividing your editing objective to more significant scene/style/subject changes and detail adding/removing and more moderate changes. Then describe the major changes in a detailed target prompt and add the more fine grained details as SEGA concepts.
+    4. **Reconstruction:** Using an empty source prompt + target prompt will lead to a perfect reconstruction
+- **Fidelity vs creativity**:
+
+    Bigger values → more fidelity, smaller values → more creativity
+
+    1. `Skip Steps`
+    2. `Warmup` (SEGA)
+    3. `Threshold`  (SEGA)
+
+    Bigger values → more creativity, smaller values → more fidelity
+
+    1. `Guidance Scale`
+    2. `Concept Guidance Scale` (SEGA)
+"""
+
+with gr.Blocks(css="style.css") as demo:
+    def update_counter(sega_concepts_counter, concept1, concept2, concept3):
+        if sega_concepts_counter == "":
+            sega_concepts_counter = sum(1 for concept in (concept1, concept2, concept3) if concept != '')
+        return sega_concepts_counter
+    def remove_concept(sega_concepts_counter, row_triggered):
+      sega_concepts_counter -= 1
+      rows_visibility = [gr.update(visible=False) for _ in range(4)]
+      
+      if(row_triggered-1 > sega_concepts_counter):
+            rows_visibility[sega_concepts_counter] = gr.update(visible=True)
+      else:
+            rows_visibility[row_triggered-1] = gr.update(visible=True)
+      
+      row1_visibility, row2_visibility, row3_visibility, row4_visibility = rows_visibility
+
+      guidance_scale_label = "Concept Guidance Scale"
+      # enable_interactive =  gr.update(interactive=True)
+      return (gr.update(visible=False),
+              gr.update(visible=False, value="",),
+              gr.update(interactive=True, value=""),
+              gr.update(visible=False,label = guidance_scale_label),
+              gr.update(interactive=True, value =False),
+              gr.update(value=DEFAULT_WARMUP_STEPS),
+              gr.update(value=DEFAULT_THRESHOLD),
+              gr.update(visible=True),
+              gr.update(interactive=True, value="custom"),
+              row1_visibility,
+              row2_visibility,
+              row3_visibility,
+              row4_visibility,
+              sega_concepts_counter
+             ) 
+    
+    
+    
+    def update_display_concept(button_label, edit_concept, neg_guidance, sega_concepts_counter):
+      sega_concepts_counter += 1
+      guidance_scale_label = "Concept Guidance Scale"
+      if(button_label=='Remove'):
+        neg_guidance = True
+        guidance_scale_label = "Negative Guidance Scale" 
+      
+      return (gr.update(visible=True), #boxn
+             gr.update(visible=True, value=edit_concept), #concept_n
+             gr.update(visible=True,label = guidance_scale_label), #guidance_scale_n
+             gr.update(value=neg_guidance),#neg_guidance_n
+             gr.update(visible=False), #row_n
+             gr.update(visible=True), #row_n+1
+             sega_concepts_counter
+             ) 
+
+
+    def display_editing_options(run_button, clear_button, sega_tab):
+      return run_button.update(visible=True), clear_button.update(visible=True), sega_tab.update(visible=True)
+    
+    def update_interactive_mode(add_button_label):
+      if add_button_label == "Clear":
+        return gr.update(interactive=False), gr.update(interactive=False)
+      else:
+        return gr.update(interactive=True), gr.update(interactive=True)
+    
+    def update_dropdown_parms(dropdown):
+        if dropdown == 'custom':
+          return DEFAULT_SEGA_CONCEPT_GUIDANCE_SCALE,DEFAULT_WARMUP_STEPS, DEFAULT_THRESHOLD
+        elif dropdown =='style':
+          return STYLE_SEGA_CONCEPT_GUIDANCE_SCALE,STYLE_WARMUP_STEPS, STYLE_THRESHOLD
+        elif dropdown =='object':
+          return OBJECT_SEGA_CONCEPT_GUIDANCE_SCALE,OBJECT_WARMUP_STEPS, OBJECT_THRESHOLD
+        elif dropdown =='faces':
+          return FACE_SEGA_CONCEPT_GUIDANCE_SCALE,FACE_WARMUP_STEPS, FACE_THRESHOLD
+
+
+    def reset_do_inversion():
+        return True
+
+    def reset_do_reconstruction():
+      do_reconstruction = True
+      return  do_reconstruction
+
+    def reset_image_caption():
+        return ""
+
+    def update_inversion_progress_visibility(input_image, do_inversion):
+      if do_inversion and not input_image is None:
+          return inversion_progress.update(visible=True)
+      else:
+        return inversion_progress.update(visible=False)
+
+    def update_edit_progress_visibility(input_image, do_inversion):
+      # if do_inversion and not input_image is None:
+      #     return inversion_progress.update(visible=True)
+      # else:
+        return inversion_progress.update(visible=True)
+
+
+    gr.HTML(intro)
+    wts = gr.State()
+    zs = gr.State()
+    reconstruction = gr.State()
+    do_inversion = gr.State(value=True)
+    do_reconstruction = gr.State(value=True)
+    sega_concepts_counter = gr.State(0)
+    image_caption = gr.State(value="")
+
+    with gr.Row():
+        input_image = gr.Image(label="Input Image", interactive=True, elem_id="input_image")
+        ddpm_edited_image = gr.Image(label=f"Pure DDPM Inversion Image", interactive=False, visible=False)
+        sega_edited_image = gr.Image(label=f"LEDITS Edited Image", interactive=False, elem_id="output_image")
+        input_image.style(height=365, width=365)
+        ddpm_edited_image.style(height=365, width=365)
+        sega_edited_image.style(height=365, width=365)
+    
+    with gr.Group(visible=False) as share_btn_container:
+        with gr.Group(elem_id="share-btn-container"):
+            community_icon = gr.HTML(community_icon_html, visible=True)
+            loading_icon = gr.HTML(loading_icon_html, visible=False)
+            share_button = gr.Button("Share to community", elem_id="share-btn", visible=True)
+        
+    with gr.Row():
+      with gr.Box(visible=False, elem_id="box1") as box1:
+        with gr.Row():
+          concept_1 = gr.Button(scale=3, value="")
+          remove_concept1 = gr.Button("x", scale=1, min_width=10)
+        with gr.Row():
+            guidnace_scale_1 = gr.Slider(label='Concept Guidance Scale', minimum=1, maximum=30,
+                            info="How strongly the concept should modify the image",
+                                                  value=DEFAULT_SEGA_CONCEPT_GUIDANCE_SCALE,
+                                                  step=0.5, interactive=True)
+      with gr.Box(visible=False, elem_id="box2") as box2:
+        with gr.Row():
+          concept_2 = gr.Button(scale=3, value="")
+          remove_concept2 = gr.Button("x", scale=1, min_width=10)
+        with gr.Row():
+          guidnace_scale_2 = gr.Slider(label='Concept Guidance Scale', minimum=1, maximum=30,
+                              info="How strongly the concept should modify the image",
+                                                    value=DEFAULT_SEGA_CONCEPT_GUIDANCE_SCALE,
+                                                    step=0.5, interactive=True)
+      with gr.Box(visible=False, elem_id="box3") as box3:
+        with gr.Row():
+          concept_3 = gr.Button(scale=3, value="")
+          remove_concept3 = gr.Button("x", scale=1, min_width=10)
+        with gr.Row():
+          guidnace_scale_3 = gr.Slider(label='Concept Guidance Scale', minimum=1, maximum=30,
+                              info="How strongly the concept should modify the image",
+                                                    value=DEFAULT_SEGA_CONCEPT_GUIDANCE_SCALE,
+                                                    step=0.5, interactive=True)
+
+
+    with gr.Row():
+        inversion_progress = gr.Textbox(visible=False, label="Inversion progress")
+        
+    with gr.Box():
+        intro_segs = gr.Markdown("Add/Remove Concepts from your Image <span style=\"font-size: 12px; color: rgb(156, 163, 175)\">with Semantic Guidance</span>")
+                  # 1st SEGA concept
+        with gr.Row().style(mobile_collapse=False) as row1:
+              with gr.Column(scale=3, min_width=100):
+                  with gr.Row().style(mobile_collapse=True):
+                      # with gr.Column(scale=3, min_width=100):
+                            edit_concept_1 = gr.Textbox(
+                                              label="Concept",
+                                              show_label=True,
+                                              max_lines=1, value="",
+                                              placeholder="E.g.: Sunglasses",
+                                          )
+                      # with gr.Column(scale=2, min_width=100):# better mobile ui
+                            dropdown1 = gr.Dropdown(label = "Edit Type", value ='custom' , choices=['custom','style', 'object', 'faces'])
+    
+
+              with gr.Column(scale=1, min_width=100, visible=False):
+                      neg_guidance_1 = gr.Checkbox(
+                          label='Remove Concept?')
+              
+              with gr.Column(scale=1, min_width=100):
+                   with gr.Row().style(mobile_collapse=False): # better mobile ui
+                       with gr.Column():
+                          add_1 = gr.Button('Add')
+                          remove_1 = gr.Button('Remove')
+             
+    
+                  # 2nd SEGA concept
+        with gr.Row(visible=False).style(equal_height=True) as row2:
+            with gr.Column(scale=3, min_width=100):
+                with gr.Row().style(mobile_collapse=True): #better mobile UI
+                    # with gr.Column(scale=3, min_width=100):
+                            edit_concept_2 = gr.Textbox(
+                                              label="Concept",
+                                              show_label=True,
+                                              max_lines=1,
+                                              placeholder="E.g.: Realistic",
+                                          )
+                    # with gr.Column(scale=2, min_width=100):# better mobile ui
+                            dropdown2 = gr.Dropdown(label = "Edit Type", value ='custom' , choices=['custom','style', 'object', 'faces'])
+
+            with gr.Column(scale=1, min_width=100, visible=False):
+                      neg_guidance_2 = gr.Checkbox(
+                          label='Remove Concept?')
+                
+            with gr.Column(scale=1, min_width=100):
+                with gr.Row().style(mobile_collapse=False): # better mobile ui
+                    with gr.Column():
+                      add_2 = gr.Button('Add')
+                      remove_2 = gr.Button('Remove')
+    
+                  # 3rd SEGA concept
+        with gr.Row(visible=False).style(equal_height=True) as row3:
+            with gr.Column(scale=3, min_width=100):
+                with gr.Row().style(mobile_collapse=True): #better mobile UI  
+                    # with gr.Column(scale=3, min_width=100):
+                            edit_concept_3 = gr.Textbox(
+                                              label="Concept",
+                                              show_label=True,
+                                              max_lines=1,
+                                              placeholder="E.g.: orange",
+                                          )
+                    # with gr.Column(scale=2, min_width=100):
+                            dropdown3 = gr.Dropdown(label = "Edit Type", value ='custom' , choices=['custom','style', 'object', 'faces'])
+            
+            with gr.Column(scale=1, min_width=100, visible=False):
+                             neg_guidance_3 = gr.Checkbox(
+                              label='Remove Concept?',visible=True)
+            
+            with gr.Column(scale=1, min_width=100):
+                with gr.Row().style(mobile_collapse=False): # better mobile ui
+                    with gr.Column():
+                         add_3 = gr.Button('Add')
+                         remove_3 = gr.Button('Remove')
+    
+        with gr.Row(visible=False).style(equal_height=True) as row4:
+            gr.Markdown("### Max of 3 concepts reached. Remove a concept to add more")
+    
+        #with gr.Row(visible=False).style(mobile_collapse=False, equal_height=True):
+        #            add_concept_button = gr.Button("+1 concept")
+
+
+    
+    with gr.Row().style(mobile_collapse=False, equal_height=True):
+                tar_prompt = gr.Textbox(
+                                label="Describe your edited image (optional)",
+                                elem_id="target_prompt",
+                                # show_label=False,
+                                max_lines=1, value="", scale=3,
+                                placeholder="Target prompt, DDPM Inversion", info = "DDPM Inversion Prompt. Can help with global changes, modify to what you would like to see"
+                            )
+                # caption_button = gr.Button("Caption Image", scale=1)
+        
+    
+    with gr.Row():
+        run_button = gr.Button("Edit your image!", visible=True)
+        
+
+    with gr.Accordion("Advanced Options", open=False):
+      with gr.Tabs() as tabs:
+
+          with gr.TabItem('General options', id=2):
+            with gr.Row():
+                with gr.Column(min_width=100):
+                   clear_button = gr.Button("Clear", visible=True)
+                   src_prompt = gr.Textbox(lines=1, label="Source Prompt", interactive=True, placeholder="")
+                   steps = gr.Number(value=100, precision=0, label="Num Diffusion Steps", interactive=True)
+                   src_cfg_scale = gr.Number(value=3.5, label=f"Source Guidance Scale", interactive=True)
+                   
+
+                with gr.Column(min_width=100):
+                    reconstruct_button = gr.Button("Show Reconstruction", visible=False)
+                    skip = gr.Slider(minimum=0, maximum=60, value=36, step=1, label="Skip Steps", interactive=True, info = "At which step to start denoising. Bigger values increase fidelity to input image")
+                    tar_cfg_scale = gr.Slider(minimum=1, maximum=30,value=15, label=f"Guidance Scale", interactive=True)
+                    seed = gr.Number(value=0, precision=0, label="Seed", interactive=True)
+                    randomize_seed = gr.Checkbox(label='Randomize seed', value=False)
+
+          with gr.TabItem('SEGA options', id=3) as sega_advanced_tab:
+             # 1st SEGA concept
+              gr.Markdown("1st concept")
+              with gr.Row().style(mobile_collapse=False, equal_height=True):
+                  warmup_1 = gr.Slider(label='Warmup', minimum=0, maximum=50,
+                                       value=DEFAULT_WARMUP_STEPS,
+                                       step=1, interactive=True, info="At which step to start applying semantic guidance. Bigger values reduce edit concept's effect")
+                  threshold_1 = gr.Slider(label='Threshold', minimum=0.5, maximum=0.99,
+                                          value=DEFAULT_THRESHOLD, step=0.01, interactive=True, 
+                                          info = "Lower the threshold for more effect (e.g. ~0.9 for style transfer)")
+
+              # 2nd SEGA concept
+              gr.Markdown("2nd concept")
+              with gr.Row() as row2_advanced:
+                  warmup_2 = gr.Slider(label='Warmup', minimum=0, maximum=50,
+                                       value=DEFAULT_WARMUP_STEPS,
+                                       step=1, interactive=True, info="At which step to start applying semantic guidance. Bigger values reduce edit concept's effect")
+                  threshold_2 = gr.Slider(label='Threshold', minimum=0.5, maximum=0.99,
+                                          value=DEFAULT_THRESHOLD,
+                                          step=0.01, interactive=True,
+                                         info = "Lower the threshold for more effect (e.g. ~0.9 for style transfer)")
+              # 3rd SEGA concept
+              gr.Markdown("3rd concept")
+              with gr.Row() as row3_advanced:
+                  warmup_3 = gr.Slider(label='Warmup', minimum=0, maximum=50,
+                                       value=DEFAULT_WARMUP_STEPS, step=1,
+                                       interactive=True, info="At which step to start applying semantic guidance. Bigger values reduce edit concept's effect")
+                  threshold_3 = gr.Slider(label='Threshold', minimum=0.5, maximum=0.99,
+                                          value=DEFAULT_THRESHOLD, step=0.01,
+                                          interactive=True,
+                                         info = "Lower the threshold for more effect (e.g. ~0.9 for style transfer)")
+
+    # caption_button.click(
+    #     fn = caption_image,
+    #     inputs = [input_image],
+    #     outputs = [tar_prompt]
+    # )
+    #neg_guidance_1.change(fn = update_label, inputs=[neg_guidance_1], outputs=[add_1])
+    #neg_guidance_2.change(fn = update_label, inputs=[neg_guidance_2], outputs=[add_2])
+    #neg_guidance_3.change(fn = update_label, inputs=[neg_guidance_3], outputs=[add_3])
+    add_1.click(fn=update_counter,
+                inputs=[sega_concepts_counter,edit_concept_1,edit_concept_2,edit_concept_3],
+                outputs=sega_concepts_counter,queue=False).then(fn = update_display_concept, inputs=[add_1, edit_concept_1, neg_guidance_1, sega_concepts_counter],  outputs=[box1, concept_1, guidnace_scale_1,neg_guidance_1,row1, row2, sega_concepts_counter],queue=False)
+    add_2.click(fn=update_counter,inputs=[sega_concepts_counter,edit_concept_1,edit_concept_2,edit_concept_3], outputs=sega_concepts_counter,queue=False).then(fn = update_display_concept, inputs=[add_2, edit_concept_2, neg_guidance_2, sega_concepts_counter],  outputs=[box2, concept_2, guidnace_scale_2,neg_guidance_2,row2, row3, sega_concepts_counter],queue=False)
+    add_3.click(fn=update_counter,inputs=[sega_concepts_counter,edit_concept_1,edit_concept_2,edit_concept_3], outputs=sega_concepts_counter,queue=False).then(fn = update_display_concept, inputs=[add_3, edit_concept_3, neg_guidance_3, sega_concepts_counter],  outputs=[box3, concept_3, guidnace_scale_3,neg_guidance_3,row3, row4, sega_concepts_counter],queue=False)
+    
+    remove_1.click(fn = update_display_concept, inputs=[remove_1, edit_concept_1, neg_guidance_1, sega_concepts_counter],  outputs=[box1, concept_1, guidnace_scale_1,neg_guidance_1,row1, row2, sega_concepts_counter],queue=False)
+    remove_2.click(fn = update_display_concept, inputs=[remove_2, edit_concept_2, neg_guidance_2 ,sega_concepts_counter],  outputs=[box2, concept_2, guidnace_scale_2,neg_guidance_2,row2, row3,sega_concepts_counter],queue=False)
+    remove_3.click(fn = update_display_concept, inputs=[remove_3, edit_concept_3, neg_guidance_3, sega_concepts_counter],  outputs=[box3, concept_3, guidnace_scale_3,neg_guidance_3, row3, row4, sega_concepts_counter],queue=False)
+    
+    remove_concept1.click(
+        fn=update_counter,inputs=[sega_concepts_counter,edit_concept_1,edit_concept_2,edit_concept_3], outputs=sega_concepts_counter,queue=False).then(
+        fn = remove_concept, inputs=[sega_concepts_counter,gr.State(1)], outputs= [box1, concept_1, edit_concept_1, guidnace_scale_1,neg_guidance_1,warmup_1, threshold_1, add_1, dropdown1, row1, row2, row3, row4, sega_concepts_counter],queue=False)
+    remove_concept2.click(
+        fn=update_counter,inputs=[sega_concepts_counter,edit_concept_1,edit_concept_2,edit_concept_3], outputs=sega_concepts_counter,queue=False).then(
+        fn = remove_concept,  inputs=[sega_concepts_counter,gr.State(2)], outputs=[box2, concept_2, edit_concept_2, guidnace_scale_2,neg_guidance_2, warmup_2, threshold_2, add_2 , dropdown2, row1, row2, row3, row4, sega_concepts_counter],queue=False)
+    remove_concept3.click(
+        fn=update_counter,inputs=[sega_concepts_counter,edit_concept_1,edit_concept_2,edit_concept_3], outputs=sega_concepts_counter,queue=False).then(
+        fn = remove_concept,inputs=[sega_concepts_counter,gr.State(3)], outputs=[box3, concept_3, edit_concept_3, guidnace_scale_3,neg_guidance_3,warmup_3, threshold_3,  add_3, dropdown3, row1, row2, row3, row4, sega_concepts_counter],queue=False)
+
+    #add_concept_button.click(fn = update_display_concept, inputs=sega_concepts_counter,
+    #           outputs= [row2, row2_advanced, row3, row3_advanced, add_concept_button, sega_concepts_counter], queue = False)
+
+    run_button.click(
+        fn=edit,
+        inputs=[input_image,
+                wts, zs,
+                tar_prompt,
+                image_caption,
+                steps,
+                skip,
+                tar_cfg_scale,
+                edit_concept_1,edit_concept_2,edit_concept_3,
+                guidnace_scale_1,guidnace_scale_2,guidnace_scale_3,
+                warmup_1, warmup_2, warmup_3,
+                neg_guidance_1, neg_guidance_2, neg_guidance_3,
+                threshold_1, threshold_2, threshold_3, do_reconstruction, reconstruction,
+                do_inversion,
+                seed, 
+                randomize_seed,
+                src_prompt,
+                src_cfg_scale
+
+
+        ],
+        outputs=[sega_edited_image, reconstruct_button, do_reconstruction, reconstruction, wts, zs, do_inversion, share_btn_container])
+    # .success(fn=update_gallery_display, inputs= [prev_output_image, sega_edited_image], outputs = [gallery, gallery, prev_output_image])
+
+
+    input_image.change(
+        fn = reset_do_inversion,
+        outputs = [do_inversion],
+        queue = False)
+    # Automatically start inverting upon input_image change
+    input_image.upload(fn = crop_image, inputs = [input_image], outputs = [input_image],queue=False).then(
+        fn = reset_do_inversion,
+        outputs = [do_inversion],
+        queue = False).then(fn = caption_image,
+        inputs = [input_image],
+        outputs = [tar_prompt, image_caption]).then(fn = update_inversion_progress_visibility, inputs =[input_image,do_inversion],
+                            outputs=[inversion_progress],queue=False).then(
+        fn=load_and_invert,
+        inputs=[input_image,
+                do_inversion,
+                seed, randomize_seed,
+                wts, zs,
+                src_prompt,
+                tar_prompt,
+                steps,
+                src_cfg_scale,
+                skip,
+                tar_cfg_scale,
+        ],
+        # outputs=[ddpm_edited_image, wts, zs, do_inversion],
+        outputs=[wts, zs, do_inversion, inversion_progress],
+    ).then(fn = update_inversion_progress_visibility, inputs =[input_image,do_inversion],
+           outputs=[inversion_progress],queue=False).then(
+              lambda: reconstruct_button.update(visible=False),
+              outputs=[reconstruct_button]).then(
+        fn = reset_do_reconstruction,
+        outputs = [do_reconstruction],
+        queue = False)
+
+
+    # Repeat inversion (and reconstruction) when these params are changed:
+    src_prompt.change(
+        fn = reset_do_inversion,
+        outputs = [do_inversion], queue = False).then(
+        fn = reset_do_reconstruction,
+        outputs = [do_reconstruction], queue = False)
+
+    steps.change(
+        fn = reset_do_inversion,
+        outputs = [do_inversion], queue = False).then(
+        fn = reset_do_reconstruction,
+        outputs = [do_reconstruction], queue = False)
+
+
+    src_cfg_scale.change(
+        fn = reset_do_inversion,
+        outputs = [do_inversion], queue = False).then(
+        fn = reset_do_reconstruction,
+        outputs = [do_reconstruction], queue = False)
+
+    # Repeat only reconstruction these params are changed:
+
+    tar_prompt.change(
+        fn = reset_do_reconstruction,
+        outputs = [do_reconstruction], queue = False)
+
+    tar_cfg_scale.change(
+        fn = reset_do_reconstruction,
+        outputs = [do_reconstruction], queue = False)
+
+    skip.change(
+        fn = reset_do_reconstruction,
+        outputs = [do_reconstruction], queue = False)
+
+    dropdown1.change(fn=update_dropdown_parms, inputs = [dropdown1], outputs = [guidnace_scale_1,warmup_1,  threshold_1], queue=False)
+    dropdown2.change(fn=update_dropdown_parms, inputs = [dropdown2], outputs = [guidnace_scale_2,warmup_2,  threshold_2], queue=False)
+    dropdown3.change(fn=update_dropdown_parms, inputs = [dropdown3], outputs = [guidnace_scale_3,warmup_3,  threshold_3], queue=False)
+
+    clear_components = [input_image,ddpm_edited_image,ddpm_edited_image,sega_edited_image, do_inversion,
+                                   src_prompt, steps, src_cfg_scale, seed,
+                                  tar_prompt, skip, tar_cfg_scale, reconstruct_button,reconstruct_button,
+                                  edit_concept_1, guidnace_scale_1,guidnace_scale_1,warmup_1,  threshold_1, neg_guidance_1,dropdown1, concept_1, concept_1, row1,
+                                  edit_concept_2, guidnace_scale_2,guidnace_scale_2,warmup_2,  threshold_2, neg_guidance_2,dropdown2, concept_2, concept_2, row2,
+                                  edit_concept_3, guidnace_scale_3,guidnace_scale_3,warmup_3,  threshold_3, neg_guidance_3,dropdown3, concept_3,concept_3, row3,
+                                  row4,sega_concepts_counter, box1, box2, box3 ]
+
+    clear_components_output_vals = [None, None,ddpm_edited_image.update(visible=False), None, True,
+                     "", DEFAULT_DIFFUSION_STEPS, DEFAULT_SOURCE_GUIDANCE_SCALE, DEFAULT_SEED,
+                     "", DEFAULT_SKIP_STEPS, DEFAULT_TARGET_GUIDANCE_SCALE, reconstruct_button.update(value="Show Reconstruction"),reconstruct_button.update(visible=False),
+                     "", DEFAULT_SEGA_CONCEPT_GUIDANCE_SCALE,guidnace_scale_1.update(visible=False), DEFAULT_WARMUP_STEPS, DEFAULT_THRESHOLD, DEFAULT_NEGATIVE_GUIDANCE, "custom","", concept_1.update(visible=False), row1.update(visible=True),
+                     "", DEFAULT_SEGA_CONCEPT_GUIDANCE_SCALE,guidnace_scale_2.update(visible=False), DEFAULT_WARMUP_STEPS, DEFAULT_THRESHOLD, DEFAULT_NEGATIVE_GUIDANCE, "custom","", concept_2.update(visible=False), row2.update(visible=False),
+                     "", DEFAULT_SEGA_CONCEPT_GUIDANCE_SCALE,guidnace_scale_3.update(visible=False), DEFAULT_WARMUP_STEPS, DEFAULT_THRESHOLD, DEFAULT_NEGATIVE_GUIDANCE, "custom","",concept_3.update(visible=False), row3.update(visible=False), row4.update(visible=False), gr.update(value=0),
+                          box1.update(visible=False), box2.update(visible=False), box3.update(visible=False)]
+
+
+    clear_button.click(lambda: clear_components_output_vals, outputs =clear_components)
+
+    reconstruct_button.click(lambda: ddpm_edited_image.update(visible=True), outputs=[ddpm_edited_image]).then(fn = reconstruct,
+                inputs = [tar_prompt,
+                image_caption,
+                tar_cfg_scale,
+                skip,
+                wts, zs,
+                do_reconstruction,
+                reconstruction,
+                          reconstruct_button],
+                outputs = [ddpm_edited_image,reconstruction, ddpm_edited_image, do_reconstruction, reconstruct_button])
+
+    randomize_seed.change(
+        fn = randomize_seed_fn,
+        inputs = [seed, randomize_seed],
+        outputs = [seed],
+        queue = False)
+
+    share_button.click(None, [], [], _js=share_js)
+    
+    gr.Examples(
+        label='Examples',
+        fn=swap_visibilities,
+        run_on_click=True,
+        examples=get_example(),
+        inputs=[input_image,
+                    edit_concept_1,
+                    edit_concept_2,
+                    tar_prompt,
+                    sega_edited_image,
+                    guidnace_scale_1,
+                    guidnace_scale_2,
+                    warmup_1,
+                    warmup_2,
+                    neg_guidance_1,
+                    neg_guidance_2,
+                    steps,
+                    skip,
+                    tar_cfg_scale,
+                    sega_concepts_counter
+               ],
+        outputs=[share_btn_container, box1, concept_1, guidnace_scale_1,neg_guidance_1, row1, row2,box2, concept_2, guidnace_scale_2,neg_guidance_2,row2, row3,sega_concepts_counter],
+        cache_examples=True
+    )
+
+
+demo.queue()
+demo.launch()

constants.py

@@ -0,0 +1,26 @@
+###############
+# conststants #
+###############
+DEFAULT_TARGET_GUIDANCE_SCALE = 15
+DEFAULT_SOURCE_GUIDANCE_SCALE = 3.5
+DEFAULT_DIFFUSION_STEPS = 100
+DEFAULT_SKIP_STEPS = 36
+DEFAULT_SEED = 0
+
+
+DEFAULT_SEGA_CONCEPT_GUIDANCE_SCALE = 7
+DEFAULT_WARMUP_STEPS = 2
+DEFAULT_THRESHOLD = 0.95
+DEFAULT_NEGATIVE_GUIDANCE=False
+
+STYLE_SEGA_CONCEPT_GUIDANCE_SCALE = 8
+STYLE_WARMUP_STEPS = 2
+STYLE_THRESHOLD = 0.92
+
+FACE_SEGA_CONCEPT_GUIDANCE_SCALE = 5
+FACE_WARMUP_STEPS = 2
+FACE_THRESHOLD = 0.95
+
+OBJECT_SEGA_CONCEPT_GUIDANCE_SCALE = 15
+OBJECT_WARMUP_STEPS = 5
+OBJECT_THRESHOLD = 0.95

inversion_utils.py

@@ -0,0 +1,275 @@
+import torch
+import os
+from tqdm import tqdm
+from PIL import Image, ImageDraw ,ImageFont
+from matplotlib import pyplot as plt
+import torchvision.transforms as T
+import os
+import yaml
+import numpy as np
+
+
+def load_512(image_path, left=0, right=0, top=0, bottom=0, device=None):
+    if type(image_path) is str:
+        image = np.array(Image.open(image_path).convert('RGB'))[:, :, :3]
+    else:
+        image = image_path
+    h, w, c = image.shape
+    left = min(left, w-1)
+    right = min(right, w - left - 1)
+    top = min(top, h - left - 1)
+    bottom = min(bottom, h - top - 1)
+    image = image[top:h-bottom, left:w-right]
+    h, w, c = image.shape
+    if h < w:
+        offset = (w - h) // 2
+        image = image[:, offset:offset + h]
+    elif w < h:
+        offset = (h - w) // 2
+        image = image[offset:offset + w]
+    image = np.array(Image.fromarray(image).resize((512, 512)))
+    image = torch.from_numpy(image).float() / 127.5 - 1
+    image = image.permute(2, 0, 1).unsqueeze(0).to(device, dtype =torch.float16)
+
+    return image 
+
+
+
+def mu_tilde(model, xt,x0, timestep):
+    "mu_tilde(x_t, x_0) DDPM paper eq. 7"
+    prev_timestep = timestep - model.scheduler.config.num_train_timesteps // model.scheduler.num_inference_steps
+    alpha_prod_t_prev = model.scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else model.scheduler.final_alpha_cumprod
+    alpha_t = model.scheduler.alphas[timestep]
+    beta_t = 1 - alpha_t 
+    alpha_bar = model.scheduler.alphas_cumprod[timestep]
+    return ((alpha_prod_t_prev ** 0.5 * beta_t) / (1-alpha_bar)) * x0 +  ((alpha_t**0.5 *(1-alpha_prod_t_prev)) / (1- alpha_bar))*xt
+
+def sample_xts_from_x0(model, x0, num_inference_steps=50):
+    """
+    Samples from P(x_1:T|x_0)
+    """
+    # torch.manual_seed(43256465436)
+    alpha_bar = model.scheduler.alphas_cumprod
+    sqrt_one_minus_alpha_bar = (1-alpha_bar) ** 0.5
+    alphas = model.scheduler.alphas
+    betas = 1 - alphas
+    variance_noise_shape = (
+            num_inference_steps,
+            model.unet.in_channels, 
+            model.unet.sample_size,
+            model.unet.sample_size)
+    
+    timesteps = model.scheduler.timesteps.to(model.device)
+    t_to_idx = {int(v):k for k,v in enumerate(timesteps)}
+    xts = torch.zeros(variance_noise_shape).to(x0.device,  dtype =torch.float16)
+    for t in reversed(timesteps):
+        idx = t_to_idx[int(t)]
+        xts[idx] = x0 * (alpha_bar[t] ** 0.5) + torch.randn_like(x0,  dtype =torch.float16) * sqrt_one_minus_alpha_bar[t]
+    xts = torch.cat([xts, x0 ],dim = 0)
+
+    return xts
+
+def encode_text(model, prompts):
+    text_input = model.tokenizer(
+        prompts,
+        padding="max_length",
+        max_length=model.tokenizer.model_max_length, 
+        truncation=True,
+        return_tensors="pt",
+    )
+    with torch.no_grad():
+        text_encoding = model.text_encoder(text_input.input_ids.to(model.device))[0]
+    return text_encoding
+
+def forward_step(model, model_output, timestep, sample):
+    next_timestep = min(model.scheduler.config.num_train_timesteps - 2,
+                        timestep + model.scheduler.config.num_train_timesteps // model.scheduler.num_inference_steps)
+
+    # 2. compute alphas, betas
+    alpha_prod_t = model.scheduler.alphas_cumprod[timestep]
+    # alpha_prod_t_next = self.scheduler.alphas_cumprod[next_timestep] if next_ltimestep >= 0 else self.scheduler.final_alpha_cumprod
+
+    beta_prod_t = 1 - alpha_prod_t
+
+    # 3. compute predicted original sample from predicted noise also called
+    # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+    pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
+
+    # 5. TODO: simple noising implementatiom
+    next_sample = model.scheduler.add_noise(pred_original_sample,
+                                    model_output,
+                                    torch.LongTensor([next_timestep]))
+    return next_sample
+
+
+def get_variance(model, timestep): #, prev_timestep):
+    prev_timestep = timestep - model.scheduler.config.num_train_timesteps // model.scheduler.num_inference_steps
+    alpha_prod_t = model.scheduler.alphas_cumprod[timestep]
+    alpha_prod_t_prev = model.scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else model.scheduler.final_alpha_cumprod
+    beta_prod_t = 1 - alpha_prod_t
+    beta_prod_t_prev = 1 - alpha_prod_t_prev
+    variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev)
+    return variance
+
+def inversion_forward_process(model, x0, 
+                            etas = None,    
+                            prog_bar = False,
+                            prompt = "",
+                            cfg_scale = 3.5,
+                            num_inference_steps=50, eps = None):
+
+    if not prompt=="":
+        text_embeddings = encode_text(model, prompt)
+    uncond_embedding = encode_text(model, "")
+    timesteps = model.scheduler.timesteps.to(model.device)
+    variance_noise_shape = (
+        num_inference_steps,
+        model.unet.in_channels, 
+        model.unet.sample_size,
+        model.unet.sample_size)
+    if etas is None or (type(etas) in [int, float] and etas == 0):
+        eta_is_zero = True
+        zs = None
+    else:
+        eta_is_zero = False
+        if type(etas) in [int, float]: etas = [etas]*model.scheduler.num_inference_steps
+        xts = sample_xts_from_x0(model, x0, num_inference_steps=num_inference_steps)
+        alpha_bar = model.scheduler.alphas_cumprod
+        zs = torch.zeros(size=variance_noise_shape, device=model.device,  dtype =torch.float16)
+        
+    t_to_idx = {int(v):k for k,v in enumerate(timesteps)}
+    xt = x0
+    op = tqdm(reversed(timesteps), desc= "Inverting...") if prog_bar else reversed(timesteps)
+
+    for t in op:
+        idx = t_to_idx[int(t)]
+        # 1. predict noise residual
+        if not eta_is_zero:
+            xt = xts[idx][None]
+                    
+        with torch.no_grad():
+            out = model.unet.forward(xt, timestep =  t, encoder_hidden_states = uncond_embedding)
+            if not prompt=="":
+                cond_out = model.unet.forward(xt, timestep=t, encoder_hidden_states = text_embeddings)
+
+        if not prompt=="":
+            ## classifier free guidance
+            noise_pred = out.sample + cfg_scale * (cond_out.sample - out.sample)
+        else:
+            noise_pred = out.sample
+
+        if eta_is_zero:
+            # 2. compute more noisy image and set x_t -> x_t+1
+            xt = forward_step(model, noise_pred, t, xt)
+
+        else: 
+            xtm1 =  xts[idx+1][None]
+            # pred of x0
+            pred_original_sample = (xt - (1-alpha_bar[t])  ** 0.5 * noise_pred ) / alpha_bar[t] ** 0.5
+            
+            # direction to xt
+            prev_timestep = t - model.scheduler.config.num_train_timesteps // model.scheduler.num_inference_steps
+            alpha_prod_t_prev = model.scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else model.scheduler.final_alpha_cumprod
+            
+            variance = get_variance(model, t)
+            pred_sample_direction = (1 - alpha_prod_t_prev - etas[idx] * variance ) ** (0.5) * noise_pred
+
+            mu_xt = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
+            
+            z = (xtm1 - mu_xt ) / ( etas[idx] * variance ** 0.5 )
+            zs[idx] = z
+
+            # correction to avoid error accumulation
+            xtm1 = mu_xt + ( etas[idx] * variance ** 0.5 )*z
+            xts[idx+1] = xtm1
+
+    if not zs is None: 
+        zs[-1] = torch.zeros_like(zs[-1]) 
+
+    return xt, zs, xts
+
+
+def reverse_step(model, model_output, timestep, sample, eta = 0, variance_noise=None):
+    # 1. get previous step value (=t-1)
+    prev_timestep = timestep - model.scheduler.config.num_train_timesteps // model.scheduler.num_inference_steps
+    # 2. compute alphas, betas
+    alpha_prod_t = model.scheduler.alphas_cumprod[timestep]
+    alpha_prod_t_prev = model.scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else model.scheduler.final_alpha_cumprod
+    beta_prod_t = 1 - alpha_prod_t
+    # 3. compute predicted original sample from predicted noise also called
+    # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+    pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
+    # 5. compute variance: "sigma_t(η)" -> see formula (16)
+    # σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1)    
+    # variance = self.scheduler._get_variance(timestep, prev_timestep)
+    variance = get_variance(model, timestep) #, prev_timestep)
+    std_dev_t = eta * variance ** (0.5)
+    # Take care of asymetric reverse process (asyrp)
+    model_output_direction = model_output
+    # 6. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+    # pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * model_output_direction
+    pred_sample_direction = (1 - alpha_prod_t_prev - eta * variance) ** (0.5) * model_output_direction
+    # 7. compute x_t without "random noise" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+    prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
+    # 8. Add noice if eta > 0
+    if eta > 0:
+        if variance_noise is None:
+            variance_noise = torch.randn(model_output.shape, device=model.device,  dtype =torch.float16)
+        sigma_z =  eta * variance ** (0.5) * variance_noise
+        prev_sample = prev_sample + sigma_z
+
+    return prev_sample
+
+def inversion_reverse_process(model,
+                    xT, 
+                    etas = 0,
+                    prompts = "",
+                    cfg_scales = None,
+                    prog_bar = False,
+                    zs = None,
+                    controller=None,
+                    asyrp = False):
+
+    batch_size = len(prompts)
+
+    cfg_scales_tensor = torch.Tensor(cfg_scales).view(-1,1,1,1).to(model.device,  dtype=torch.float16)
+
+    text_embeddings = encode_text(model, prompts)
+    uncond_embedding = encode_text(model, [""] * batch_size)
+
+    if etas is None: etas = 0
+    if type(etas) in [int, float]: etas = [etas]*model.scheduler.num_inference_steps
+    assert len(etas) == model.scheduler.num_inference_steps
+    timesteps = model.scheduler.timesteps.to(model.device)
+
+    xt = xT.expand(batch_size, -1, -1, -1)
+    op = tqdm(timesteps[-zs.shape[0]:]) if prog_bar else timesteps[-zs.shape[0]:] 
+
+    t_to_idx = {int(v):k for k,v in enumerate(timesteps[-zs.shape[0]:])}
+
+    for t in op:
+        idx = t_to_idx[int(t)]        
+        ## Unconditional embedding
+        with torch.no_grad():
+            uncond_out = model.unet.forward(xt, timestep =  t, 
+                                            encoder_hidden_states = uncond_embedding)
+
+            ## Conditional embedding  
+        if prompts:  
+            with torch.no_grad():
+                cond_out = model.unet.forward(xt, timestep =  t, 
+                                                encoder_hidden_states = text_embeddings)
+            
+        
+        z = zs[idx] if not zs is None else None
+        z = z.expand(batch_size, -1, -1, -1)
+        if prompts:
+            ## classifier free guidance
+            noise_pred = uncond_out.sample + cfg_scales_tensor * (cond_out.sample - uncond_out.sample)
+        else: 
+            noise_pred = uncond_out.sample
+        # 2. compute less noisy image and set x_t -> x_t-1  
+        xt = reverse_step(model, noise_pred, t, xt, eta = etas[idx], variance_noise = z) 
+        if controller is not None:
+            xt = controller.step_callback(xt)        
+    return xt, zs

modified_pipeline_semantic_stable_diffusion.py

@@ -0,0 +1,753 @@
+
+import inspect
+import warnings
+from itertools import repeat
+from typing import Callable, List, Optional, Union
+
+import torch
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
+
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.models import AutoencoderKL, UNet2DConditionModel
+from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+from diffusers.schedulers import KarrasDiffusionSchedulers
+from diffusers.utils import logging, randn_tensor
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
+# from . import SemanticStableDiffusionPipelineOutput
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class SemanticStableDiffusionPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for text-to-image generation with latent editing.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    This model builds on the implementation of ['StableDiffusionPipeline']
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`CLIPTextModel`]):
+            Frozen text-encoder. Stable Diffusion uses the text portion of
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latens. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+        safety_checker ([`Q16SafetyChecker`]):
+            Classification module that estimates whether generated images could be considered offensive or harmful.
+            Please, refer to the [model card](https://huggingface.co/CompVis/stable-diffusion-v1-4) for details.
+        feature_extractor ([`CLIPImageProcessor`]):
+            Model that extracts features from generated images to be used as inputs for the `safety_checker`.
+    """
+
+    _optional_components = ["safety_checker", "feature_extractor"]
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        tokenizer: CLIPTokenizer,
+        unet: UNet2DConditionModel,
+        scheduler: KarrasDiffusionSchedulers,
+        safety_checker: StableDiffusionSafetyChecker,
+        feature_extractor: CLIPImageProcessor,
+        requires_safety_checker: bool = True,
+    ):
+        super().__init__()
+
+        if safety_checker is None and requires_safety_checker:
+            logger.warning(
+                f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+                " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+                " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+                " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+                " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+                " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+            )
+
+        if safety_checker is not None and feature_extractor is None:
+            raise ValueError(
+                "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+                " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+            )
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            unet=unet,
+            scheduler=scheduler,
+            safety_checker=safety_checker,
+            feature_extractor=feature_extractor,
+        )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+        self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+    def run_safety_checker(self, image, device, dtype):
+        if self.safety_checker is None:
+            has_nsfw_concept = None
+        else:
+            if torch.is_tensor(image):
+                feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+            else:
+                feature_extractor_input = self.image_processor.numpy_to_pil(image)
+            safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+            image, has_nsfw_concept = self.safety_checker(
+                images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+            )
+        return image, has_nsfw_concept
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+    def decode_latents(self, latents):
+        warnings.warn(
+            "The decode_latents method is deprecated and will be removed in a future version. Please"
+            " use VaeImageProcessor instead",
+            FutureWarning,
+        )
+        latents = 1 / self.vae.config.scaling_factor * latents
+        image = self.vae.decode(latents, return_dict=False)[0]
+        image = (image / 2 + 0.5).clamp(0, 1)
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+        return image
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.check_inputs
+    def check_inputs(
+        self,
+        prompt,
+        height,
+        width,
+        callback_steps,
+        negative_prompt=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+    ):
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if (callback_steps is None) or (
+            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+        ):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+    def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+        shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: Union[str, List[str]],
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 7.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: int = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: int = 1,
+        editing_prompt: Optional[Union[str, List[str]]] = None,
+        editing_prompt_embeddings: Optional[torch.Tensor] = None,
+        reverse_editing_direction: Optional[Union[bool, List[bool]]] = False,
+        edit_guidance_scale: Optional[Union[float, List[float]]] = 5,
+        edit_warmup_steps: Optional[Union[int, List[int]]] = 10,
+        edit_cooldown_steps: Optional[Union[int, List[int]]] = None,
+        edit_threshold: Optional[Union[float, List[float]]] = 0.9,
+        edit_momentum_scale: Optional[float] = 0.1,
+        edit_mom_beta: Optional[float] = 0.4,
+        edit_weights: Optional[List[float]] = None,
+        sem_guidance: Optional[List[torch.Tensor]] = None,
+
+        # DDPM additions
+        use_ddpm: bool = False,
+        wts: Optional[List[torch.Tensor]] = None,
+        zs: Optional[List[torch.Tensor]] = None
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`):
+                The prompt or prompts to guide the image generation.
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The width in pixels of the generated image.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+                if `guidance_scale` is less than `1`).
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+            callback (`Callable`, *optional*):
+                A function that will be called every `callback_steps` steps during inference. The function will be
+                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+            callback_steps (`int`, *optional*, defaults to 1):
+                The frequency at which the `callback` function will be called. If not specified, the callback will be
+                called at every step.
+            editing_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to use for Semantic guidance. Semantic guidance is disabled by setting
+                `editing_prompt = None`. Guidance direction of prompt should be specified via
+                `reverse_editing_direction`.
+            editing_prompt_embeddings (`torch.Tensor>`, *optional*):
+                Pre-computed embeddings to use for semantic guidance. Guidance direction of embedding should be
+                specified via `reverse_editing_direction`.
+            reverse_editing_direction (`bool` or `List[bool]`, *optional*, defaults to `False`):
+                Whether the corresponding prompt in `editing_prompt` should be increased or decreased.
+            edit_guidance_scale (`float` or `List[float]`, *optional*, defaults to 5):
+                Guidance scale for semantic guidance. If provided as list values should correspond to `editing_prompt`.
+                `edit_guidance_scale` is defined as `s_e` of equation 6 of [SEGA
+                Paper](https://arxiv.org/pdf/2301.12247.pdf).
+            edit_warmup_steps (`float` or `List[float]`, *optional*, defaults to 10):
+                Number of diffusion steps (for each prompt) for which semantic guidance will not be applied. Momentum
+                will still be calculated for those steps and applied once all warmup periods are over.
+                `edit_warmup_steps` is defined as `delta` (δ) of [SEGA Paper](https://arxiv.org/pdf/2301.12247.pdf).
+            edit_cooldown_steps (`float` or `List[float]`, *optional*, defaults to `None`):
+                Number of diffusion steps (for each prompt) after which semantic guidance will no longer be applied.
+            edit_threshold (`float` or `List[float]`, *optional*, defaults to 0.9):
+                Threshold of semantic guidance.
+            edit_momentum_scale (`float`, *optional*, defaults to 0.1):
+                Scale of the momentum to be added to the semantic guidance at each diffusion step. If set to 0.0
+                momentum will be disabled. Momentum is already built up during warmup, i.e. for diffusion steps smaller
+                than `sld_warmup_steps`. Momentum will only be added to latent guidance once all warmup periods are
+                finished. `edit_momentum_scale` is defined as `s_m` of equation 7 of [SEGA
+                Paper](https://arxiv.org/pdf/2301.12247.pdf).
+            edit_mom_beta (`float`, *optional*, defaults to 0.4):
+                Defines how semantic guidance momentum builds up. `edit_mom_beta` indicates how much of the previous
+                momentum will be kept. Momentum is already built up during warmup, i.e. for diffusion steps smaller
+                than `edit_warmup_steps`. `edit_mom_beta` is defined as `beta_m` (β) of equation 8 of [SEGA
+                Paper](https://arxiv.org/pdf/2301.12247.pdf).
+            edit_weights (`List[float]`, *optional*, defaults to `None`):
+                Indicates how much each individual concept should influence the overall guidance. If no weights are
+                provided all concepts are applied equally. `edit_mom_beta` is defined as `g_i` of equation 9 of [SEGA
+                Paper](https://arxiv.org/pdf/2301.12247.pdf).
+            sem_guidance (`List[torch.Tensor]`, *optional*):
+                List of pre-generated guidance vectors to be applied at generation. Length of the list has to
+                correspond to `num_inference_steps`.
+
+        Returns:
+            [`~pipelines.semantic_stable_diffusion.SemanticStableDiffusionPipelineOutput`] or `tuple`:
+            [`~pipelines.semantic_stable_diffusion.SemanticStableDiffusionPipelineOutput`] if `return_dict` is True,
+            otherwise a `tuple. When returning a tuple, the first element is a list with the generated images, and the
+            second element is a list of `bool`s denoting whether the corresponding generated image likely represents
+            "not-safe-for-work" (nsfw) content, according to the `safety_checker`.
+        """
+        # 0. Default height and width to unet
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(prompt, height, width, callback_steps)
+
+        # 2. Define call parameters
+        batch_size = 1 if isinstance(prompt, str) else len(prompt)
+
+        if editing_prompt:
+            enable_edit_guidance = True
+            if isinstance(editing_prompt, str):
+                editing_prompt = [editing_prompt]
+            enabled_editing_prompts = len(editing_prompt)
+        elif editing_prompt_embeddings is not None:
+            enable_edit_guidance = True
+            enabled_editing_prompts = editing_prompt_embeddings.shape[0]
+        else:
+            enabled_editing_prompts = 0
+            enable_edit_guidance = False
+
+        # get prompt text embeddings
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=self.tokenizer.model_max_length,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+
+        if text_input_ids.shape[-1] > self.tokenizer.model_max_length:
+            removed_text = self.tokenizer.batch_decode(text_input_ids[:, self.tokenizer.model_max_length :])
+            logger.warning(
+                "The following part of your input was truncated because CLIP can only handle sequences up to"
+                f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+            )
+            text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length]
+        text_embeddings = self.text_encoder(text_input_ids.to(self.device))[0]
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        bs_embed, seq_len, _ = text_embeddings.shape
+        text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1)
+        text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+        if enable_edit_guidance:
+            # get safety text embeddings
+            if editing_prompt_embeddings is None:
+                edit_concepts_input = self.tokenizer(
+                    [x for item in editing_prompt for x in repeat(item, batch_size)],
+                    padding="max_length",
+                    max_length=self.tokenizer.model_max_length,
+                    return_tensors="pt",
+                )
+
+                edit_concepts_input_ids = edit_concepts_input.input_ids
+
+                if edit_concepts_input_ids.shape[-1] > self.tokenizer.model_max_length:
+                    removed_text = self.tokenizer.batch_decode(
+                        edit_concepts_input_ids[:, self.tokenizer.model_max_length :]
+                    )
+                    logger.warning(
+                        "The following part of your input was truncated because CLIP can only handle sequences up to"
+                        f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+                    )
+                    edit_concepts_input_ids = edit_concepts_input_ids[:, : self.tokenizer.model_max_length]
+                edit_concepts = self.text_encoder(edit_concepts_input_ids.to(self.device))[0]
+            else:
+                edit_concepts = editing_prompt_embeddings.to(self.device).repeat(batch_size, 1, 1)
+
+            # duplicate text embeddings for each generation per prompt, using mps friendly method
+            bs_embed_edit, seq_len_edit, _ = edit_concepts.shape
+            edit_concepts = edit_concepts.repeat(1, num_images_per_prompt, 1)
+            edit_concepts = edit_concepts.view(bs_embed_edit * num_images_per_prompt, seq_len_edit, -1)
+
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+        # get unconditional embeddings for classifier free guidance
+
+        if do_classifier_free_guidance:
+            uncond_tokens: List[str]
+            if negative_prompt is None:
+                uncond_tokens = [""]
+            elif type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif isinstance(negative_prompt, str):
+                uncond_tokens = [negative_prompt]
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_tokens = negative_prompt
+
+            max_length = text_input_ids.shape[-1]
+            uncond_input = self.tokenizer(
+                uncond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+            uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]
+
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = uncond_embeddings.shape[1]
+            uncond_embeddings = uncond_embeddings.repeat(batch_size, num_images_per_prompt, 1)
+            uncond_embeddings = uncond_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and text embeddings into a single batch
+            # to avoid doing two forward passes
+            if enable_edit_guidance:
+                text_embeddings = torch.cat([uncond_embeddings, text_embeddings, edit_concepts])
+            else:
+                text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
+        # get the initial random noise unless the user supplied it
+
+        # 4. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=self.device)
+        timesteps = self.scheduler.timesteps
+        if use_ddpm:
+          t_to_idx = {int(v):k for k,v in enumerate(timesteps[-zs.shape[0]:])}
+          timesteps = timesteps[-zs.shape[0]:] 
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.unet.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            text_embeddings.dtype,
+            self.device,
+            generator,
+            latents,
+        )
+
+        # 6. Prepare extra step kwargs.
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # Initialize edit_momentum to None
+        edit_momentum = None
+
+        self.uncond_estimates = None
+        self.text_estimates = None
+        self.edit_estimates = None
+        self.sem_guidance = None
+
+        for i, t in enumerate(self.progress_bar(timesteps)):
+            # expand the latents if we are doing classifier free guidance
+            latent_model_input = (
+                torch.cat([latents] * (2 + enabled_editing_prompts)) if do_classifier_free_guidance else latents
+            )
+            latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+            # predict the noise residual
+            noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample
+
+            # perform guidance
+            if do_classifier_free_guidance:
+                noise_pred_out = noise_pred.chunk(2 + enabled_editing_prompts)  # [b,4, 64, 64]
+                noise_pred_uncond, noise_pred_text = noise_pred_out[0], noise_pred_out[1]
+                noise_pred_edit_concepts = noise_pred_out[2:]
+
+                # default text guidance
+                noise_guidance = guidance_scale * (noise_pred_text - noise_pred_uncond)
+                # noise_guidance = (noise_pred_text - noise_pred_edit_concepts[0])
+
+                if self.uncond_estimates is None:
+                    self.uncond_estimates = torch.zeros((num_inference_steps + 1, *noise_pred_uncond.shape))
+                self.uncond_estimates[i] = noise_pred_uncond.detach().cpu()
+
+                if self.text_estimates is None:
+                    self.text_estimates = torch.zeros((num_inference_steps + 1, *noise_pred_text.shape))
+                self.text_estimates[i] = noise_pred_text.detach().cpu()
+
+                if self.edit_estimates is None and enable_edit_guidance:
+                    self.edit_estimates = torch.zeros(
+                        (num_inference_steps + 1, len(noise_pred_edit_concepts), *noise_pred_edit_concepts[0].shape)
+                    )
+
+                if self.sem_guidance is None:
+                    self.sem_guidance = torch.zeros((num_inference_steps + 1, *noise_pred_text.shape))
+
+                if edit_momentum is None:
+                    edit_momentum = torch.zeros_like(noise_guidance)
+
+                if enable_edit_guidance:
+                    concept_weights = torch.zeros(
+                        (len(noise_pred_edit_concepts), noise_guidance.shape[0]),
+                        device=self.device,
+                        dtype=noise_guidance.dtype,
+                    )
+                    noise_guidance_edit = torch.zeros(
+                        (len(noise_pred_edit_concepts), *noise_guidance.shape),
+                        device=self.device,
+                        dtype=noise_guidance.dtype,
+                    )
+                    # noise_guidance_edit = torch.zeros_like(noise_guidance)
+                    warmup_inds = []
+                    for c, noise_pred_edit_concept in enumerate(noise_pred_edit_concepts):
+                        self.edit_estimates[i, c] = noise_pred_edit_concept
+                        if isinstance(edit_guidance_scale, list):
+                            edit_guidance_scale_c = edit_guidance_scale[c]
+                        else:
+                            edit_guidance_scale_c = edit_guidance_scale
+
+                        if isinstance(edit_threshold, list):
+                            edit_threshold_c = edit_threshold[c]
+                        else:
+                            edit_threshold_c = edit_threshold
+                        if isinstance(reverse_editing_direction, list):
+                            reverse_editing_direction_c = reverse_editing_direction[c]
+                        else:
+                            reverse_editing_direction_c = reverse_editing_direction
+                        if edit_weights:
+                            edit_weight_c = edit_weights[c]
+                        else:
+                            edit_weight_c = 1.0
+                        if isinstance(edit_warmup_steps, list):
+                            edit_warmup_steps_c = edit_warmup_steps[c]
+                        else:
+                            edit_warmup_steps_c = edit_warmup_steps
+
+                        if isinstance(edit_cooldown_steps, list):
+                            edit_cooldown_steps_c = edit_cooldown_steps[c]
+                        elif edit_cooldown_steps is None:
+                            edit_cooldown_steps_c = i + 1
+                        else:
+                            edit_cooldown_steps_c = edit_cooldown_steps
+                        if i >= edit_warmup_steps_c:
+                            warmup_inds.append(c)
+                        if i >= edit_cooldown_steps_c:
+                            noise_guidance_edit[c, :, :, :, :] = torch.zeros_like(noise_pred_edit_concept)
+                            continue
+
+                        noise_guidance_edit_tmp = noise_pred_edit_concept - noise_pred_uncond
+                        # tmp_weights = (noise_pred_text - noise_pred_edit_concept).sum(dim=(1, 2, 3))
+                        tmp_weights = (noise_guidance - noise_pred_edit_concept).sum(dim=(1, 2, 3))
+
+                        tmp_weights = torch.full_like(tmp_weights, edit_weight_c)  # * (1 / enabled_editing_prompts)
+                        if reverse_editing_direction_c:
+                            noise_guidance_edit_tmp = noise_guidance_edit_tmp * -1
+                        concept_weights[c, :] = tmp_weights
+
+                        noise_guidance_edit_tmp = noise_guidance_edit_tmp * edit_guidance_scale_c
+
+                        # torch.quantile function expects float32
+                        if noise_guidance_edit_tmp.dtype == torch.float32:
+                            tmp = torch.quantile(
+                                torch.abs(noise_guidance_edit_tmp).flatten(start_dim=2),
+                                edit_threshold_c,
+                                dim=2,
+                                keepdim=False,
+                            )
+                        else:
+                            tmp = torch.quantile(
+                                torch.abs(noise_guidance_edit_tmp).flatten(start_dim=2).to(torch.float32),
+                                edit_threshold_c,
+                                dim=2,
+                                keepdim=False,
+                            ).to(noise_guidance_edit_tmp.dtype)
+
+                        noise_guidance_edit_tmp = torch.where(
+                            torch.abs(noise_guidance_edit_tmp) >= tmp[:, :, None, None],
+                            noise_guidance_edit_tmp,
+                            torch.zeros_like(noise_guidance_edit_tmp),
+                        )
+                        noise_guidance_edit[c, :, :, :, :] = noise_guidance_edit_tmp
+
+                        # noise_guidance_edit = noise_guidance_edit + noise_guidance_edit_tmp
+
+                    warmup_inds = torch.tensor(warmup_inds).to(self.device)
+                    if len(noise_pred_edit_concepts) > warmup_inds.shape[0] > 0:
+                        concept_weights = concept_weights.to("cpu")  # Offload to cpu
+                        noise_guidance_edit = noise_guidance_edit.to("cpu")
+
+                        concept_weights_tmp = torch.index_select(concept_weights.to(self.device), 0, warmup_inds)
+                        concept_weights_tmp = torch.where(
+                            concept_weights_tmp < 0, torch.zeros_like(concept_weights_tmp), concept_weights_tmp
+                        )
+                        concept_weights_tmp = concept_weights_tmp / concept_weights_tmp.sum(dim=0)
+                        # concept_weights_tmp = torch.nan_to_num(concept_weights_tmp)
+
+                        noise_guidance_edit_tmp = torch.index_select(
+                            noise_guidance_edit.to(self.device), 0, warmup_inds
+                        )
+                        noise_guidance_edit_tmp = torch.einsum(
+                            "cb,cbijk->bijk", concept_weights_tmp, noise_guidance_edit_tmp
+                        )
+                        noise_guidance_edit_tmp = noise_guidance_edit_tmp
+                        noise_guidance = noise_guidance + noise_guidance_edit_tmp
+
+                        self.sem_guidance[i] = noise_guidance_edit_tmp.detach().cpu()
+
+                        del noise_guidance_edit_tmp
+                        del concept_weights_tmp
+                        concept_weights = concept_weights.to(self.device)
+                        noise_guidance_edit = noise_guidance_edit.to(self.device)
+
+                    concept_weights = torch.where(
+                        concept_weights < 0, torch.zeros_like(concept_weights), concept_weights
+                    )
+
+                    concept_weights = torch.nan_to_num(concept_weights)
+
+                    noise_guidance_edit = torch.einsum("cb,cbijk->bijk", concept_weights, noise_guidance_edit)
+
+                    noise_guidance_edit = noise_guidance_edit + edit_momentum_scale * edit_momentum
+
+                    edit_momentum = edit_mom_beta * edit_momentum + (1 - edit_mom_beta) * noise_guidance_edit
+
+                    if warmup_inds.shape[0] == len(noise_pred_edit_concepts):
+                        noise_guidance = noise_guidance + noise_guidance_edit
+                        self.sem_guidance[i] = noise_guidance_edit.detach().cpu()
+
+                if sem_guidance is not None:
+                    edit_guidance = sem_guidance[i].to(self.device)
+                    noise_guidance = noise_guidance + edit_guidance
+
+                noise_pred = noise_pred_uncond + noise_guidance
+            ## ddpm ###########################################################
+            if use_ddpm:
+              
+              idx = t_to_idx[int(t)]
+              z = zs[idx] if not zs is None else None
+
+              # 1. get previous step value (=t-1)
+              prev_timestep = t - self.scheduler.config.num_train_timesteps // self.scheduler.num_inference_steps
+              # 2. compute alphas, betas
+              alpha_prod_t = self.scheduler.alphas_cumprod[t]
+              alpha_prod_t_prev = self.scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.scheduler.final_alpha_cumprod
+              beta_prod_t = 1 - alpha_prod_t
+              
+              # 3. compute predicted original sample from predicted noise also called
+              # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+              pred_original_sample = (latents - beta_prod_t ** (0.5) * noise_pred) / alpha_prod_t ** (0.5)
+              
+              
+              # 5. compute variance: "sigma_t(η)" -> see formula (16)
+              # σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1)    
+              # variance = self.scheduler._get_variance(timestep, prev_timestep)
+              # variance = get_variance(model, t) #, prev_timestep)
+              beta_prod_t_prev = 1 - alpha_prod_t_prev
+              variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev)
+              
+            
+              
+              std_dev_t = eta * variance ** (0.5)
+              # Take care of asymetric reverse process (asyrp)
+              noise_pred_direction = noise_pred
+              
+              # 6. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+              # pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * model_output_direction
+              pred_sample_direction = (1 - alpha_prod_t_prev - eta * variance) ** (0.5) * noise_pred_direction
+              
+              # 7. compute x_t without "random noise" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+              prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
+              # 8. Add noice if eta > 0
+              if eta > 0:
+                  if z is None:
+                      z = torch.randn(noise_pred.shape, device=self.device)
+                  sigma_z =  eta * variance ** (0.5) * z
+                  latents = prev_sample + sigma_z
+
+            ## ddpm ##########################################################
+                # compute the previous noisy sample x_t -> x_t-1
+            else: #if not use_ddpm:
+              latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+            # call the callback, if provided
+            if callback is not None and i % callback_steps == 0:
+                callback(i, t, latents)
+
+        
+        #  # 8. Post-processing
+        # image = self.decode_latents(latents)
+
+        # # 9. Run safety checker
+        # image, has_nsfw_concept = self.run_safety_checker(image, self.device, text_embeddings.dtype)
+
+        # # 10. Convert to PIL
+        # if output_type == "pil":
+        #     image = self.numpy_to_pil(image)
+
+        # if not return_dict:
+        #     return (image, has_nsfw_concept)
+
+        # return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
+        
+        # 8. Post-processing
+        if not output_type == "latent":
+            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+            image, has_nsfw_concept = self.run_safety_checker(image, self.device, text_embeddings.dtype)
+        else:
+            image = latents
+            has_nsfw_concept = None
+
+        if has_nsfw_concept is None:
+            do_denormalize = [True] * image.shape[0]
+        else:
+            do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+        image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+        if not return_dict:
+            return (image, has_nsfw_concept)
+
+        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)

requirements.txt

@@ -0,0 +1,5 @@
+diffusers
+accelerate
+transformers
+torch
+torchvision

share_btn.py

@@ -0,0 +1,109 @@
+community_icon_html = """<svg id="share-btn-share-icon" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32">
+    <path d="M20.6081 3C21.7684 3 22.8053 3.49196 23.5284 4.38415C23.9756 4.93678 24.4428 5.82749 24.4808 7.16133C24.9674 7.01707 25.4353 6.93643 25.8725 6.93643C26.9833 6.93643 27.9865 7.37587 28.696 8.17411C29.6075 9.19872 30.0124 10.4579 29.8361 11.7177C29.7523 12.3177 29.5581 12.8555 29.2678 13.3534C29.8798 13.8646 30.3306 14.5763 30.5485 15.4322C30.719 16.1032 30.8939 17.5006 29.9808 18.9403C30.0389 19.0342 30.0934 19.1319 30.1442 19.2318C30.6932 20.3074 30.7283 21.5229 30.2439 22.6548C29.5093 24.3704 27.6841 25.7219 24.1397 27.1727C21.9347 28.0753 19.9174 28.6523 19.8994 28.6575C16.9842 29.4379 14.3477 29.8345 12.0653 29.8345C7.87017 29.8345 4.8668 28.508 3.13831 25.8921C0.356375 21.6797 0.754104 17.8269 4.35369 14.1131C6.34591 12.058 7.67023 9.02782 7.94613 8.36275C8.50224 6.39343 9.97271 4.20438 12.4172 4.20438H12.4179C12.6236 4.20438 12.8314 4.2214 13.0364 4.25468C14.107 4.42854 15.0428 5.06476 15.7115 6.02205C16.4331 5.09583 17.134 4.359 17.7682 3.94323C18.7242 3.31737 19.6794 3 20.6081 3ZM20.6081 5.95917C20.2427 5.95917 19.7963 6.1197 19.3039 6.44225C17.7754 7.44319 14.8258 12.6772 13.7458 14.7131C13.3839 15.3952 12.7655 15.6837 12.2086 15.6837C11.1036 15.6837 10.2408 14.5497 12.1076 13.1085C14.9146 10.9402 13.9299 7.39584 12.5898 7.1776C12.5311 7.16799 12.4731 7.16355 12.4172 7.16355C11.1989 7.16355 10.6615 9.33114 10.6615 9.33114C10.6615 9.33114 9.0863 13.4148 6.38031 16.206C3.67434 18.998 3.5346 21.2388 5.50675 24.2246C6.85185 26.2606 9.42666 26.8753 12.0653 26.8753C14.8021 26.8753 17.6077 26.2139 19.1799 25.793C19.2574 25.7723 28.8193 22.984 27.6081 20.6107C27.4046 20.212 27.0693 20.0522 26.6471 20.0522C24.9416 20.0522 21.8393 22.6726 20.5057 22.6726C20.2076 22.6726 19.9976 22.5416 19.9116 22.222C19.3433 20.1173 28.552 19.2325 27.7758 16.1839C27.639 15.6445 27.2677 15.4256 26.746 15.4263C24.4923 15.4263 19.4358 19.5181 18.3759 19.5181C18.2949 19.5181 18.2368 19.4937 18.2053 19.4419C17.6743 18.557 17.9653 17.9394 21.7082 15.6009C25.4511 13.2617 28.0783 11.8545 26.5841 10.1752C26.4121 9.98141 26.1684 9.8956 25.8725 9.8956C23.6001 9.89634 18.2311 14.9403 18.2311 14.9403C18.2311 14.9403 16.7821 16.496 15.9057 16.496C15.7043 16.496 15.533 16.4139 15.4169 16.2112C14.7956 15.1296 21.1879 10.1286 21.5484 8.06535C21.7928 6.66715 21.3771 5.95917 20.6081 5.95917Z" fill="#FF9D00"></path>
+    <path d="M5.50686 24.2246C3.53472 21.2387 3.67446 18.9979 6.38043 16.206C9.08641 13.4147 10.6615 9.33111 10.6615 9.33111C10.6615 9.33111 11.2499 6.95933 12.59 7.17757C13.93 7.39581 14.9139 10.9401 12.1069 13.1084C9.29997 15.276 12.6659 16.7489 13.7459 14.713C14.8258 12.6772 17.7747 7.44316 19.304 6.44221C20.8326 5.44128 21.9089 6.00204 21.5484 8.06532C21.188 10.1286 14.795 15.1295 15.4171 16.2118C16.0391 17.2934 18.2312 14.9402 18.2312 14.9402C18.2312 14.9402 25.0907 8.49588 26.5842 10.1752C28.0776 11.8545 25.4512 13.2616 21.7082 15.6008C17.9646 17.9393 17.6744 18.557 18.2054 19.4418C18.7372 20.3266 26.9998 13.1351 27.7759 16.1838C28.5513 19.2324 19.3434 20.1173 19.9117 22.2219C20.48 24.3274 26.3979 18.2382 27.6082 20.6107C28.8193 22.9839 19.2574 25.7722 19.18 25.7929C16.0914 26.62 8.24723 28.3726 5.50686 24.2246Z" fill="#FFD21E"></path>
+</svg>"""
+
+loading_icon_html = """<svg id="share-btn-loading-icon" style="display:none;" class="animate-spin"
+   style="color: #ffffff; 
+"
+   xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" fill="none" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 24 24"><circle style="opacity: 0.25;" cx="12" cy="12" r="10" stroke="white" stroke-width="4"></circle><path style="opacity: 0.75;" fill="white" d="M4 12a8 8 0 018-8V0C5.373 0 0 5.373 0 12h4zm2 5.291A7.962 7.962 0 014 12H0c0 3.042 1.135 5.824 3 7.938l3-2.647z"></path></svg>"""
+
+share_js = r"""async () => {
+    async function uploadFile(file){
+        const UPLOAD_URL = 'https://huggingface.co/uploads';
+        const response = await fetch(UPLOAD_URL, {
+            method: 'POST',
+            headers: {
+                'Content-Type': file.type,
+                'X-Requested-With': 'XMLHttpRequest',
+            },
+            body: file,
+        });
+        const url = await response.text();
+        return url;
+    }
+
+    function getButtonText(componentId) {
+        const buttonEl = gradioEl.querySelector(`${componentId} button`);
+        return buttonEl ? buttonEl.textContent : '';
+    }
+
+    const gradioEl = document.querySelector('body > gradio-app');
+    const imgEls = [gradioEl.querySelector('#input_image img'), gradioEl.querySelector('#output_image img')];
+    const concepts = [
+        { value: getButtonText('#box1'), parent: gradioEl.querySelector('#box1 span[data-testid="block-info"]') },
+        { value: getButtonText('#box2'), parent: gradioEl.querySelector('#box2 span[data-testid="block-info"]') },
+        { value: getButtonText('#box3'), parent: gradioEl.querySelector('#box3 span[data-testid="block-info"]') }
+    ];
+
+    const promptTxt = gradioEl.querySelector('#target_prompt input').value;
+    const shareBtnEl = gradioEl.querySelector('#share-btn');
+    const shareIconEl = gradioEl.querySelector('#share-btn-share-icon');
+    const loadingIconEl = gradioEl.querySelector('#share-btn-loading-icon');
+    if(!imgEls[1]){
+        return;
+    };
+    shareBtnEl.style.pointerEvents = 'none';
+    shareIconEl.style.display = 'none';
+    loadingIconEl.style.removeProperty('display');
+
+    async function processImage(imgEl, imgId) {
+        const res = await fetch(imgEl.src);
+        const blob = await res.blob();
+        const fileType = blob.type.includes('png') ? 'png' : 'jpg';
+        const fileName = `diffuse-the-rest-${imgId}.${fileType}`;
+        return new File([blob], fileName, { type: blob.type });
+    }
+
+    const files = await Promise.all(imgEls.map((imgEl, index) => processImage(imgEl, Date.now() + index % 200)));
+    const urls = await Promise.all(files.map((file) => uploadFile(file)));
+
+    const labels = ['Source image', 'Target image'];
+    const htmlImgs = urls.map((url, index) => `<div>${labels[index]}: <img src='${url}' width='400' height='400' style="margin: 0"></div>`);
+
+    let descriptionMd = `<div style='display: flex; flex-wrap: wrap; column-gap: 0.75rem;'>${htmlImgs.join(`\n`)}</div>`;
+
+    if (promptTxt) {
+        descriptionMd += `<span style="font-size:1.2em">Target image prompt: <b>${promptTxt}</b></span><br>`;
+    } else {
+        descriptionMd += `<span style="font-size:1.2em">Target image prompt: <b>""</b></span><br>`;
+    }
+
+    const conceptHeaders = [];
+    const conceptDescriptions = [];
+    const conceptTableRows = [];
+    concepts.forEach((concept, index) => {
+        if (concept.value) {
+            const label = concept.parent.textContent.includes('Negative') ? `remove concept` : `add concept`;
+            conceptHeaders.push(`<th>${label}</th>`);
+            conceptDescriptions.push(`${label}: ${concept.value}`);
+            conceptTableRows.push(`<td>${concept.value}</td>`);
+        }
+    });
+
+    let title = 'Editing';
+    if (promptTxt) {
+        title += ` "${promptTxt}"`;
+    }
+    if (conceptDescriptions.length > 0) {
+        title += ` to ${conceptDescriptions.join(', ')}`;
+        descriptionMd += `<table style="font-size: 1.1em">
+    <tr>
+    ${conceptHeaders.join('\n')}
+    </tr>
+    <tr>
+    ${conceptTableRows.join('\n')}
+    </tr>
+    </table>`;
+    }
+
+    const params = new URLSearchParams({
+        title: title,
+        description: descriptionMd,
+    });
+    const paramsStr = params.toString();
+    window.open(`https://huggingface.co/spaces/editing-images/ledits/discussions/new?${paramsStr}`, '_blank');
+    shareBtnEl.style.removeProperty('pointer-events');
+    shareIconEl.style.removeProperty('display');
+    loadingIconEl.style.display = 'none';
+}"""

style.css

@@ -0,0 +1,93 @@
+/*
+This CSS file is modified from:
+https://huggingface.co/spaces/DeepFloyd/IF/blob/main/style.css
+*/
+
+h1 {
+  text-align: center;
+}
+
+.gradio-container {
+  font-family: 'IBM Plex Sans', sans-serif;
+}
+
+.gr-button {
+  color: white;
+  border-color: black;
+  background: black;
+}
+
+input[type='range'] {
+  accent-color: black;
+}
+
+.dark input[type='range'] {
+  accent-color: #dfdfdf;
+}
+
+.container {
+  max-width: 730px;
+  margin: auto;
+  padding-top: 1.5rem;
+}
+
+
+.gr-button:focus {
+  border-color: rgb(147 197 253 / var(--tw-border-opacity));
+  outline: none;
+  box-shadow: var(--tw-ring-offset-shadow), var(--tw-ring-shadow), var(--tw-shadow, 0 0 #0000);
+  --tw-border-opacity: 1;
+  --tw-ring-offset-shadow: var(--tw-ring-inset) 0 0 0 var(--tw-ring-offset-width) var(--tw-ring-offset-color);
+  --tw-ring-shadow: var(--tw-ring-inset) 0 0 0 calc(3px var(--tw-ring-offset-width)) var(--tw-ring-color);
+  --tw-ring-color: rgb(191 219 254 / var(--tw-ring-opacity));
+  --tw-ring-opacity: .5;
+}
+
+.gr-form {
+  flex: 1 1 50%;
+  border-top-right-radius: 0;
+  border-bottom-right-radius: 0;
+}
+
+#prompt-container {
+  gap: 0;
+}
+
+#prompt-text-input,
+#negative-prompt-text-input {
+  padding: .45rem 0.625rem
+}
+
+/* #component-16 {
+  border-top-width: 1px !important;
+  margin-top: 1em
+} */
+
+.image_duplication {
+  position: absolute;
+  width: 100px;
+  left: 50px
+}
+
+#component-0 {
+  max-width: 730px;
+  margin: auto;
+  padding-top: 1.5rem;
+}
+
+#share-btn-container {
+    display: flex; padding-left: 0.5rem !important; padding-right: 0.5rem !important; background-color: #000000; justify-content: center; align-items: center; border-radius: 9999px !important; width: 13rem; margin-left: auto;
+}
+#share-btn {
+    all: initial; color: #ffffff;font-weight: 600; cursor:pointer; font-family: 'IBM Plex Sans', sans-serif; margin-left: 0.5rem !important; padding-top: 0.25rem !important; padding-bottom: 0.25rem !important;
+}
+#share-btn * {
+    all: unset;
+}
+#share-btn-container div:nth-child(-n+2){
+    width: auto !important;
+    min-height: 0px !important;
+}
+#share-btn-container .wrap {
+    display: none !important;
+}

utils.py

@@ -0,0 +1,114 @@
+import PIL
+from PIL import Image, ImageDraw ,ImageFont
+from matplotlib import pyplot as plt
+import torchvision.transforms as T
+import os
+import torch
+import yaml
+
+def show_torch_img(img):
+    img = to_np_image(img)
+    plt.imshow(img)
+    plt.axis("off")
+
+def to_np_image(all_images):
+    all_images = (all_images.permute(0, 2, 3, 1) * 127.5 + 128).clamp(0, 255).to(torch.uint8).cpu().numpy()[0]
+    return all_images
+
+def tensor_to_pil(tensor_imgs):
+    if type(tensor_imgs) == list:
+        tensor_imgs = torch.cat(tensor_imgs)
+    tensor_imgs = (tensor_imgs / 2 + 0.5).clamp(0, 1)
+    to_pil = T.ToPILImage()
+    pil_imgs = [to_pil(img) for img in tensor_imgs]
+    return pil_imgs
+
+def pil_to_tensor(pil_imgs):
+    to_torch = T.ToTensor()
+    if type(pil_imgs) == PIL.Image.Image:
+        tensor_imgs = to_torch(pil_imgs).unsqueeze(0)*2-1
+    elif type(pil_imgs) == list:
+        tensor_imgs = torch.cat([to_torch(pil_imgs).unsqueeze(0)*2-1 for img in pil_imgs]).to(device)
+    else:
+        raise Exception("Input need to be PIL.Image or list of PIL.Image")
+    return tensor_imgs
+
+
+## TODO implement this
+# n = 10
+# num_rows = 4
+# num_col = n // num_rows
+# num_col  = num_col + 1 if n % num_rows else num_col
+# num_col
+def add_margin(pil_img, top = 0, right = 0, bottom = 0,
+                    left = 0, color = (255,255,255)):
+    width, height = pil_img.size
+    new_width = width + right + left
+    new_height = height + top + bottom
+    result = Image.new(pil_img.mode, (new_width, new_height), color)
+
+    result.paste(pil_img, (left, top))
+    return result
+
+def image_grid(imgs, rows = 1, cols = None,
+                    size = None,
+                   titles = None, text_pos = (0, 0)):
+    if type(imgs) == list and type(imgs[0]) == torch.Tensor:
+        imgs = torch.cat(imgs)
+    if type(imgs) == torch.Tensor:
+        imgs = tensor_to_pil(imgs)
+
+    if not size is None:
+        imgs = [img.resize((size,size)) for img in imgs]
+    if cols is None:
+        cols = len(imgs)
+    assert len(imgs) >= rows*cols
+
+    top=20
+    w, h = imgs[0].size
+    delta = 0
+    if len(imgs)> 1 and not imgs[1].size[1] == h:
+        delta = top
+        h = imgs[1].size[1]
+    if not titles is  None:
+        font = ImageFont.truetype("/usr/share/fonts/truetype/freefont/FreeMono.ttf",
+                                    size = 20, encoding="unic")
+        h = top + h
+    grid = Image.new('RGB', size=(cols*w, rows*h+delta))
+    for i, img in enumerate(imgs):
+
+        if not titles is  None:
+            img = add_margin(img, top = top, bottom = 0,left=0)
+            draw = ImageDraw.Draw(img)
+            draw.text(text_pos, titles[i],(0,0,0),
+            font = font)
+        if not delta == 0 and i > 0:
+           grid.paste(img, box=(i%cols*w, i//cols*h+delta))
+        else:
+            grid.paste(img, box=(i%cols*w, i//cols*h))
+
+    return grid
+
+
+"""
+input_folder - dataset folder
+"""
+def load_dataset(input_folder):
+    # full_file_names = glob.glob(input_folder)
+    # class_names = [x[0] for x in os.walk(input_folder)]
+    class_names = next(os.walk(input_folder))[1]
+    class_names[:] = [d for d in class_names if not d[0] == '.']
+    file_names=[]
+    for class_name in class_names:
+        cur_path = os.path.join(input_folder, class_name)
+        filenames = next(os.walk(cur_path), (None, None, []))[2]
+        filenames = [f for f in filenames if not f[0] == '.']
+        file_names.append(filenames)
+    return class_names, file_names
+
+
+def dataset_from_yaml(yaml_location):
+    with open(yaml_location, 'r') as stream:
+        data_loaded = yaml.safe_load(stream)
+
+    return data_loaded