local inference enhancements

app.py

@@ -1,26 +1,57 @@
 import gradio as gr
 import numpy as np
 import random
+import os
+from pathlib import Path
 
 # import spaces #[uncomment to use ZeroGPU]
-from diffusers import DiffusionPipeline
+from diffusers import DiffusionPipeline, StableDiffusionPipeline, schedulers
 import torch
 
+MODEL_REPO_ID = os.environ.get('MODEL_REPO_ID', 'myxlmynx/cyberrealistic_classic40')
+MODEL_REPO_LOCAL = os.environ.get('MODEL_REPO_LOCAL', '')
+MODEL_REPO_NAME = os.environ.get('MODEL_REPO_NAME', 'CyberRealistic Classic 4.0')
+
 device = "cuda" if torch.cuda.is_available() else "cpu"
-model_repo_id = "myxlmynx/cyberrealistic_classic40"  # Replace to the model you would like to use
+print("Running on " + device)
 
 if torch.cuda.is_available():
     torch_dtype = torch.float16
 else:
     torch_dtype = torch.float32
 
-pipe = DiffusionPipeline.from_pretrained(model_repo_id, torch_dtype=torch_dtype)
+print("Loading " + MODEL_REPO_ID)
+if MODEL_REPO_LOCAL and Path(MODEL_REPO_LOCAL).is_file():
+    pipe = StableDiffusionPipeline.from_single_file(MODEL_REPO_LOCAL, torch_dtype=torch_dtype)
+else:
+    pipe = DiffusionPipeline.from_pretrained(MODEL_REPO_ID, torch_dtype=torch_dtype)
+
+extra_inference_parameters = {}
+
+# add accel LoRA to boost generation speed
+pipe.load_lora_weights("wangfuyun/PCM_Weights",
+    subfolder='sd15', weight_name='pcm_sd15_smallcfg_2step_converted.safetensors',
+    adapter_name='pcm_smallcfg_2step')
+pipe.set_adapters(['pcm_smallcfg_2step'], adapter_weights=[1.0])
+pipe.fuse_lora()
+
+# for very low step counts with PCM
+#pipe.scheduler = schedulers.DDIMScheduler(timestep_spacing='trailing',
+#  clip_sample=False, set_alpha_to_one=False)
+pipe.scheduler = schedulers.TCDScheduler()
+extra_inference_parameters['eta'] = 0.3
+#pipe.scheduler = schedulers.LCMScheduler()
+#pipe.scheduler = schedulers.EulerAncestralDiscreteScheduler()
+
+# lib default will fry the image
+default_guidance_scale = 1
+
 pipe = pipe.to(device)
 
 MAX_SEED = np.iinfo(np.int32).max
+MIN_IMAGE_SIZE = 128
 MAX_IMAGE_SIZE = 1024
 
-
 # @spaces.GPU #[uncomment to use ZeroGPU]
 def infer(
     prompt,
@@ -38,6 +69,9 @@ def infer(
 
     generator = torch.Generator().manual_seed(seed)
 
+    if guidance_scale == 0:
+        guidance_scale = default_guidance_scale
+
     image = pipe(
         prompt=prompt,
         negative_prompt=negative_prompt,
@@ -46,6 +80,7 @@ def infer(
         width=width,
         height=height,
         generator=generator,
+        **extra_inference_parameters
     ).images[0]
 
     return image, seed
@@ -64,9 +99,14 @@ css = """
 }
 """
 
-with gr.Blocks(css=css) as demo_cpu:
+with gr.Blocks(css=css) as demo_device:
     with gr.Column(elem_id="col-container"):
-        gr.Markdown(" # Text-to-Image Gradio Template")
+        gr.Markdown("# " + MODEL_REPO_NAME + " - on " + device.upper())
+
+        if device == 'cpu':
+            gr.Markdown("Note: running on CPU, generation will be very slow. Expect at least" +
+              " a minute for minimal parameters (512x512 image, guidance <= 1, <=4 steps).\n" +
+              "It's also on a single queue, so clone this space for experimenting with it.")
 
         with gr.Row():
             prompt = gr.Text(
@@ -102,18 +142,18 @@ with gr.Blocks(css=css) as demo_cpu:
             with gr.Row():
                 width = gr.Slider(
                     label="Width",
-                    minimum=256,
+                    minimum=MIN_IMAGE_SIZE,
                     maximum=MAX_IMAGE_SIZE,
                     step=32,
-                    value=1024,  # Replace with defaults that work for your model
+                    value=512,
                 )
 
                 height = gr.Slider(
                     label="Height",
-                    minimum=256,
+                    minimum=MIN_IMAGE_SIZE,
                     maximum=MAX_IMAGE_SIZE,
                     step=32,
-                    value=1024,  # Replace with defaults that work for your model
+                    value=768,
                 )
 
             with gr.Row():
@@ -122,7 +162,7 @@ with gr.Blocks(css=css) as demo_cpu:
                     minimum=0.0,
                     maximum=10.0,
                     step=0.1,
-                    value=0.0,  # Replace with defaults that work for your model
+                    value=0.0,
                 )
 
                 num_inference_steps = gr.Slider(
@@ -130,7 +170,7 @@ with gr.Blocks(css=css) as demo_cpu:
                     minimum=1,
                     maximum=50,
                     step=1,
-                    value=2,  # Replace with defaults that work for your model
+                    value=3,
                 )
 
         gr.Examples(examples=examples, inputs=[prompt])
@@ -151,9 +191,9 @@ with gr.Blocks(css=css) as demo_cpu:
     )
 
 
-demo_inference = gr.load(model_repo_id, src='models')
+demo_inference = gr.load(MODEL_REPO_ID, title=MODEL_REPO_NAME, src='models')
 
-demo = gr.TabbedInterface([demo_inference, demo_cpu], ["Inference API", "CPU"])
+demo = gr.TabbedInterface([demo_inference, demo_device], ["Inference API", device.upper()])
 
 if __name__ == "__main__":
     demo.launch()