import gradio as gr
import numpy as np
import random
import os
import spaces #[uncomment to use ZeroGPU]
from diffusers import DiffusionPipeline
import torch
from diffusers import DiffusionPipeline, UNet2DConditionModel, LCMScheduler, AutoencoderTiny, FluxPipeline
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
import sys
sys.path.append('.')
from utils.lora import LoRANetwork, DEFAULT_TARGET_REPLACE, UNET_TARGET_REPLACE_MODULE_CONV
# Model configurations
SDXL_CONCEPTS = [
"alien", "ancient ruins", "animal", "bike", "car", "Citadel",
"coral", "cowboy", "face", "futuristic cities", "monster",
"mystical creature", "planet", "plant", "robot", "sculpture",
"spaceship", "statue", "studio", "video game", "wizard"
]
FLUX_CONCEPTS = [
"alien",
"ancient ruins",
"animal",
"bike",
"car",
"Citadel",
"face",
"futuristic cities",
"mystical creature",
"planet",
"plant",
"robot",
"spaceship",
"statue",
"studio",
"video game",
"wizard"
]
model_repo_id = "stabilityai/stable-diffusion-xl-base-1.0"
repo_name = "tianweiy/DMD2"
ckpt_name = "dmd2_sdxl_4step_unet_fp16.bin"
device = "cuda" if torch.cuda.is_available() else "cpu"
if torch.cuda.is_available():
torch_dtype = torch.bfloat16
else:
torch_dtype = torch.float32
# Load model.
unet = UNet2DConditionModel.from_config(model_repo_id, subfolder="unet").to(device, torch_dtype)
unet.load_state_dict(torch.load(hf_hub_download(repo_name, ckpt_name)))
pipe = DiffusionPipeline.from_pretrained(model_repo_id, unet=unet, torch_dtype=torch_dtype).to(device)
pipe.scheduler = LCMScheduler.from_config(pipe.scheduler.config)
pipe.vae = AutoencoderTiny.from_pretrained("madebyollin/taesdxl", torch_dtype=torch_dtype).to(device)
unet = pipe.unet
## Change these parameters based on how you trained your sliderspace sliders
train_method = 'xattn-strict'
rank = 1
alpha =1
networks = {}
modules = DEFAULT_TARGET_REPLACE
modules += UNET_TARGET_REPLACE_MODULE_CONV
for i in range(1):
networks[i] = LoRANetwork(
unet,
rank=int(rank),
multiplier=1.0,
alpha=int(alpha),
train_method=train_method,
fast_init=True,
).to(device, dtype=torch_dtype)
MAX_SEED = np.iinfo(np.int32).max
MAX_IMAGE_SIZE = 1024
base_model_id = "black-forest-labs/FLUX.1-schnell"
max_sequence_length = 256
flux_pipe = FluxPipeline.from_pretrained(base_model_id, torch_dtype=torch_dtype)
flux_pipe.vae = AutoencoderTiny.from_pretrained("madebyollin/taef1", torch_dtype=torch_dtype)
flux_pipe = flux_pipe.to(device)
# pipe.enable_sequential_cpu_offload()
transformer = flux_pipe.transformer
## Change these parameters based on how you trained your sliderspace sliders
train_method = 'flux-attn'
rank = 1
alpha =1
flux_networks = {}
modules = DEFAULT_TARGET_REPLACE
modules += UNET_TARGET_REPLACE_MODULE_CONV
for i in range(1):
flux_networks[i] = LoRANetwork(
transformer,
rank=int(rank),
multiplier=1.0,
alpha=int(alpha),
train_method=train_method,
fast_init=True,
).to(device, dtype=torch_dtype)
def update_sliderspace_choices(model_choice):
return gr.Dropdown(
choices=SDXL_CONCEPTS if model_choice == "SDXL-DMD" else FLUX_CONCEPTS,
label="SliderSpace Concept",
value=SDXL_CONCEPTS[0] if model_choice == "SDXL-DMD" else FLUX_CONCEPTS[0]
)
@spaces.GPU #[uncomment to use ZeroGPU]
def infer(
prompt,
negative_prompt,
seed,
randomize_seed,
width,
height,
guidance_scale,
num_inference_steps,
slider_space,
discovered_directions,
slider_scale,
model_choice,
progress=gr.Progress(track_tqdm=True),
):
if randomize_seed:
seed = random.randint(0, MAX_SEED)
if model_choice == 'SDXL-DMD':
sliderspace_path = f"sliderspace_weights/{slider_space}/slider_{int(discovered_directions.split(' ')[-1])-1}.pt"
for net in networks:
networks[net].load_state_dict(torch.load(sliderspace_path))
networks[net].set_lora_slider(slider_scale)
with networks[0]:
pass
# original image
generator = torch.Generator().manual_seed(seed)
image = pipe(
prompt=prompt,
negative_prompt=negative_prompt,
guidance_scale=guidance_scale,
num_inference_steps=num_inference_steps,
width=width,
height=height,
generator=generator,
).images[0]
# edited image
generator = torch.Generator().manual_seed(seed)
with networks[0]:
slider_image = pipe(
prompt=prompt,
negative_prompt=negative_prompt,
guidance_scale=guidance_scale,
num_inference_steps=num_inference_steps,
width=width,
height=height,
generator=generator,
).images[0]
else:
sliderspace_path = f"flux_sliderspace_weights/{slider_space}/slider_{int(discovered_directions.split(' ')[-1])-1}.pt"
for net in flux_networks:
flux_networks[net].load_state_dict(torch.load(sliderspace_path))
flux_networks[net].set_lora_slider(slider_scale)
with flux_networks[0]:
pass
# original image
generator = torch.Generator().manual_seed(seed)
image = flux_pipe(
prompt=prompt,
guidance_scale=guidance_scale,
num_inference_steps=num_inference_steps,
width=width,
height=height,
generator=generator,
max_sequence_length = 256,
).images[0]
# edited image
generator = torch.Generator().manual_seed(seed)
with flux_networks[0]:
slider_image = flux_pipe(
prompt=prompt,
guidance_scale=guidance_scale,
num_inference_steps=num_inference_steps,
width=width,
height=height,
generator=generator,
max_sequence_length = 256,
).images[0]
return image, slider_image, seed
examples = [
"Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
"An astronaut riding a green horse",
"A delicious ceviche cheesecake slice",
]
css = """
#col-container {
margin: 0 auto;
max-width: 640px;
}
"""
ORIGINAL_SPACE_ID = 'baulab/SliderSpace'
SPACE_ID = os.getenv('SPACE_ID')
SHARED_UI_WARNING = f'''## You can duplicate and use it with a gpu with at least 24GB, or clone this repository to run on your own machine.
'''
with gr.Blocks(css=css) as demo:
with gr.Column(elem_id="col-container"):
gr.Markdown(" # SliderSpace: Decomposing Visual Capabilities of Diffusion Models")
# Adding links under the title
gr.Markdown("""
🔗 [Project Page](https://sliderspace.baulab.info) |
💻 [GitHub Code](https://github.com/rohitgandikota/sliderspace)
""")
with gr.Row():
prompt = gr.Text(
label="Prompt",
show_label=False,
max_lines=1,
placeholder="Enter your prompt",
container=False,
)
run_button = gr.Button("Run", scale=0, variant="primary")
# Add model selection dropdown
model_choice = gr.Dropdown(
choices=["SDXL-DMD", "FLUX-Schnell"],
label="Model",
value="SDXL-DMD"
)
# New dropdowns side by side
with gr.Row():
slider_space = gr.Dropdown(
choices=SDXL_CONCEPTS,
label="SliderSpace Concept",
value=SDXL_CONCEPTS[0]
)
discovered_directions = gr.Dropdown(
choices=[f"Slider {i}" for i in range(1, 11)],
label="Discovered Directions",
value="Slider 1"
)
slider_scale = gr.Slider(
label="Slider Scale",
minimum=-4,
maximum=4,
step=0.1,
value=1,
)
with gr.Row():
result = gr.Image(label="Original Image", show_label=True)
slider_result = gr.Image(label="Discovered Edit Direction", show_label=True)
with gr.Accordion("Advanced Settings", open=False):
negative_prompt = gr.Text(
label="Negative prompt",
max_lines=1,
placeholder="Enter a negative prompt",
visible=False,
)
seed = gr.Slider(
label="Seed",
minimum=0,
maximum=MAX_SEED,
step=1,
value=0,
)
randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
with gr.Row():
width = gr.Slider(
label="Width",
minimum=256,
maximum=MAX_IMAGE_SIZE,
step=32,
value=1024, # Replace with defaults that work for your model
)
height = gr.Slider(
label="Height",
minimum=256,
maximum=MAX_IMAGE_SIZE,
step=32,
value=1024, # Replace with defaults that work for your model
)
with gr.Row():
guidance_scale = gr.Slider(
label="Guidance scale",
minimum=0.0,
maximum=2.0,
step=0.1,
value=0.0, # Replace with defaults that work for your model
)
num_inference_steps = gr.Slider(
label="Number of inference steps",
minimum=1,
maximum=50,
step=1,
value=4, # Replace with defaults that work for your model
)
# Add event handler for model selection
model_choice.change(
fn=update_sliderspace_choices,
inputs=[model_choice],
outputs=[slider_space]
)
# gr.Examples(examples=examples, inputs=[prompt])
gr.on(
triggers=[run_button.click, prompt.submit],
fn=infer,
inputs=[
prompt,
negative_prompt,
seed,
randomize_seed,
width,
height,
guidance_scale,
num_inference_steps,
slider_space,
discovered_directions,
slider_scale,
model_choice
],
outputs=[result, slider_result, seed],
)
if __name__ == "__main__":
demo.launch(share=True)
# import gradio as gr
# import numpy as np
# import random
# import os
# import spaces #[uncomment to use ZeroGPU]
# from diffusers import DiffusionPipeline
# import torch
# from diffusers import DiffusionPipeline, UNet2DConditionModel, LCMScheduler
# from huggingface_hub import hf_hub_download
# from safetensors.torch import load_file
# import sys
# sys.path.append('.')
# from utils.lora import LoRANetwork, DEFAULT_TARGET_REPLACE, UNET_TARGET_REPLACE_MODULE_CONV
# model_repo_id = "stabilityai/stable-diffusion-xl-base-1.0"
# repo_name = "tianweiy/DMD2"
# ckpt_name = "dmd2_sdxl_4step_unet_fp16.bin"
# device = "cuda" if torch.cuda.is_available() else "cpu"
# if torch.cuda.is_available():
# torch_dtype = torch.bfloat16
# else:
# torch_dtype = torch.float32
# # Load model.
# unet = UNet2DConditionModel.from_config(model_repo_id, subfolder="unet").to(device, torch_dtype)
# unet.load_state_dict(torch.load(hf_hub_download(repo_name, ckpt_name)))
# pipe = DiffusionPipeline.from_pretrained(model_repo_id, unet=unet, torch_dtype=torch_dtype).to(device)
# pipe.scheduler = LCMScheduler.from_config(pipe.scheduler.config)
# unet = pipe.unet
# ## Change these parameters based on how you trained your sliderspace sliders
# train_method = 'xattn-strict'
# rank = 1
# alpha =1
# networks = {}
# modules = DEFAULT_TARGET_REPLACE
# modules += UNET_TARGET_REPLACE_MODULE_CONV
# for i in range(1):
# networks[i] = LoRANetwork(
# unet,
# rank=int(rank),
# multiplier=1.0,
# alpha=int(alpha),
# train_method=train_method,
# fast_init=True,
# ).to(device, dtype=torch_dtype)
# MAX_SEED = np.iinfo(np.int32).max
# MAX_IMAGE_SIZE = 1024
# @spaces.GPU #[uncomment to use ZeroGPU]
# def infer(
# prompt,
# negative_prompt,
# seed,
# randomize_seed,
# width,
# height,
# guidance_scale,
# num_inference_steps,
# slider_space,
# discovered_directions,
# slider_scale,
# progress=gr.Progress(track_tqdm=True),
# ):
# if randomize_seed:
# seed = random.randint(0, MAX_SEED)
# sliderspace_path = f"sliderspace_weights/{slider_space}/slider_{int(discovered_directions.split(' ')[-1])-1}.pt"
# for net in networks:
# networks[net].load_state_dict(torch.load(sliderspace_path))
# for net in networks:
# networks[net].set_lora_slider(slider_scale)
# with networks[0]:
# pass
# # original image
# generator = torch.Generator().manual_seed(seed)
# image = pipe(
# prompt=prompt,
# negative_prompt=negative_prompt,
# guidance_scale=guidance_scale,
# num_inference_steps=num_inference_steps,
# width=width,
# height=height,
# generator=generator,
# ).images[0]
# # edited image
# generator = torch.Generator().manual_seed(seed)
# with networks[0]:
# slider_image = pipe(
# prompt=prompt,
# negative_prompt=negative_prompt,
# guidance_scale=guidance_scale,
# num_inference_steps=num_inference_steps,
# width=width,
# height=height,
# generator=generator,
# ).images[0]
# return image, slider_image, seed
# examples = [
# "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
# "An astronaut riding a green horse",
# "A delicious ceviche cheesecake slice",
# ]
# css = """
# #col-container {
# margin: 0 auto;
# max-width: 640px;
# }
# """
# ORIGINAL_SPACE_ID = 'baulab/SliderSpace'
# SPACE_ID = os.getenv('SPACE_ID')
# SHARED_UI_WARNING = f'''## You can duplicate and use it with a gpu with at least 24GB, or clone this repository to run on your own machine.
#
# '''
# with gr.Blocks(css=css) as demo:
# with gr.Column(elem_id="col-container"):
# gr.Markdown(" # SliderSpace: Decomposing Visual Capabilities of Diffusion Models")
# # Adding links under the title
# gr.Markdown("""
# 🔗 [Project Page](https://sliderspace.baulab.info) |
# 💻 [GitHub Code](https://github.com/rohitgandikota/sliderspace)
# """)
# with gr.Row():
# prompt = gr.Text(
# label="Prompt",
# show_label=False,
# max_lines=1,
# placeholder="Enter your prompt",
# container=False,
# )
# run_button = gr.Button("Run", scale=0, variant="primary")
# # New dropdowns side by side
# with gr.Row():
# slider_space = gr.Dropdown(
# choices= [
# "alien",
# "ancient ruins",
# "animal",
# "bike",
# "car",
# "Citadel",
# "coral",
# "cowboy",
# "face",
# "futuristic cities",
# "monster",
# "mystical creature",
# "planet",
# "plant",
# "robot",
# "sculpture",
# "spaceship",
# "statue",
# "studio",
# "video game",
# "wizard"
# ],
# label="SliderSpace",
# value="spaceship"
# )
# discovered_directions = gr.Dropdown(
# choices=[f"Slider {i}" for i in range(1, 11)],
# label="Discovered Directions",
# value="Slider 1"
# )
# slider_scale = gr.Slider(
# label="Slider Scale",
# minimum=-4,
# maximum=4,
# step=0.1,
# value=1,
# )
# with gr.Row():
# result = gr.Image(label="Original Image", show_label=True)
# slider_result = gr.Image(label="Discovered Edit Direction", show_label=True)
# with gr.Accordion("Advanced Settings", open=False):
# negative_prompt = gr.Text(
# label="Negative prompt",
# max_lines=1,
# placeholder="Enter a negative prompt",
# visible=False,
# )
# seed = gr.Slider(
# label="Seed",
# minimum=0,
# maximum=MAX_SEED,
# step=1,
# value=0,
# )
# randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
# with gr.Row():
# width = gr.Slider(
# label="Width",
# minimum=256,
# maximum=MAX_IMAGE_SIZE,
# step=32,
# value=1024, # Replace with defaults that work for your model
# )
# height = gr.Slider(
# label="Height",
# minimum=256,
# maximum=MAX_IMAGE_SIZE,
# step=32,
# value=1024, # Replace with defaults that work for your model
# )
# with gr.Row():
# guidance_scale = gr.Slider(
# label="Guidance scale",
# minimum=0.0,
# maximum=2.0,
# step=0.1,
# value=0.0, # Replace with defaults that work for your model
# )
# num_inference_steps = gr.Slider(
# label="Number of inference steps",
# minimum=1,
# maximum=50,
# step=1,
# value=4, # Replace with defaults that work for your model
# )
# # gr.Examples(examples=examples, inputs=[prompt])
# gr.on(
# triggers=[run_button.click, prompt.submit],
# fn=infer,
# inputs=[
# prompt,
# negative_prompt,
# seed,
# randomize_seed,
# width,
# height,
# guidance_scale,
# num_inference_steps,
# slider_space,
# discovered_directions,
# slider_scale
# ],
# outputs=[result, slider_result, seed],
# )
# if __name__ == "__main__":
# demo.launch(share=True)