Support SDXL-Lightning and fix some errors for baseline. (#1)

- Support SDXL-Lightning and fix some errors for baseline. (e30a6e165c3b3aef3a743d749a57db76f4c8405c)


Co-authored-by: Jiaxiang Cheng <jiaxiangc@users.noreply.huggingface.co>

app.py

@@ -2,25 +2,35 @@
 
 
 import os
+
 os.system("pip install -U peft")
 import random
 
 import gradio as gr
 import numpy as np
 import PIL.Image
+
 import spaces
 import torch
-from diffusers import AutoPipelineForText2Image, DPMSolverMultistepScheduler
+from diffusers import (
+    StableDiffusionXLPipeline,
+    UNet2DConditionModel,
+    EulerDiscreteScheduler,
+)
 from huggingface_hub import hf_hub_download
-from diffusers.models.attention_processor import AttnProcessor2_0
+from safetensors.torch import load_file
 
 DESCRIPTION = """
 # Res-Adapter :Domain Consistent Resolution Adapter for Diffusion Models
 **Demo by [ameer azam] - [Twitter](https://twitter.com/Ameerazam18) - [GitHub](https://github.com/AMEERAZAM08)) - [Hugging Face](https://huggingface.co/ameerazam08)**
-This is a demo of  https://huggingface.co/jiaxiangc/res-adapter LORAs by ByteDance
+This is a demo of  https://huggingface.co/jiaxiangc/res-adapter ResAdapter by ByteDance.
+
+ByteDance provide a demo of [ResAdapter](https://huggingface.co/jiaxiangc/res-adapter) with [SDXL-Lightning-Step4](https://huggingface.co/ByteDance/SDXL-Lightning) to expand resolution range from 1024-only to 256~1024.
 """
 if not torch.cuda.is_available():
-    DESCRIPTION += "\n<h1>Running on CPU 🥶 This demo does not work on CPU.</a> instead</h1>"
+    DESCRIPTION += (
+        "\n<h1>Running on CPU 🥶 This demo does not work on CPU.</a> instead</h1>"
+    )
 
 MAX_SEED = np.iinfo(np.int32).max
 CACHE_EXAMPLES = torch.cuda.is_available() and os.getenv("CACHE_EXAMPLES", "0") == "1"
@@ -29,21 +39,26 @@ USE_TORCH_COMPILE = os.getenv("USE_TORCH_COMPILE") == "1"
 ENABLE_CPU_OFFLOAD = os.getenv("ENABLE_CPU_OFFLOAD") == "1"
 
 device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
-pipe = AutoPipelineForText2Image.from_pretrained('stabilityai/stable-diffusion-xl-base-1.0',use_safetensors=True)# torch_dtype=torch.float16, variant="safetensors")
-pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config, use_karras_sigmas=True, algorithm_type="sde-dpmsolver++")
 
+base = "stabilityai/stable-diffusion-xl-base-1.0"
+repo = "ByteDance/SDXL-Lightning"
+ckpt = "sdxl_lightning_4step_unet.safetensors"  # Use the correct ckpt for your step setting!
 
+unet = UNet2DConditionModel.from_config(base, subfolder="unet").to("cuda", torch.float16)
+unet.load_state_dict(load_file(hf_hub_download(repo, ckpt), device="cuda"))
+pipe = StableDiffusionXLPipeline.from_pretrained(base, unet=unet, torch_dtype=torch.float16, variant="fp16")
+pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing")
+pipe = pipe.to(device)
 
+# Load resadapter
 pipe.load_lora_weights(
     hf_hub_download(
-        repo_id="jiaxiangc/res-adapter", 
-        subfolder="sdxl-i", 
+        repo_id="jiaxiangc/res-adapter",
+        subfolder="sdxl-i",
         filename="resolution_lora.safetensors",
     ),
     adapter_name="res_adapter",
 )
-pipe.set_adapters(["res_adapter"], adapter_weights=[1.0])
-pipe = pipe.to(device)
 
 def randomize_seed_fn(seed: int, randomize_seed: bool) -> int:
     if randomize_seed:
@@ -63,11 +78,11 @@ def generate(
     seed: int = 0,
     width: int = 1024,
     height: int = 1024,
-    guidance_scale_base: float = 5.0,
-    num_inference_steps_base: int = 20,
+    guidance_scale: float = 0,
+    num_inference_steps: int = 4,
     progress=gr.Progress(track_tqdm=True),
 ) -> PIL.Image.Image:
-    print(f"** Generating image for: \"{prompt}\" **")
+    print(f'** Generating image for: "{prompt}" **')
     generator = torch.Generator().manual_seed(seed)
 
     if not use_negative_prompt:
@@ -76,46 +91,51 @@ def generate(
         prompt_2 = None  # type: ignore
     if not use_negative_prompt_2:
         negative_prompt_2 = None  # type: ignore
-    res_adapt=pipe(
+
+    pipe.set_adapters(["res_adapter"], adapter_weights=[0.0])
+    base_image = pipe(
         prompt=prompt,
         negative_prompt=negative_prompt,
         prompt_2=prompt_2,
         negative_prompt_2=negative_prompt_2,
         width=width,
         height=height,
-        guidance_scale=guidance_scale_base,
-        num_inference_steps=num_inference_steps_base,
-        generator=generator,
+        num_inference_steps=num_inference_steps,
+        guidance_scale=guidance_scale,
         output_type="pil",
+        generator=generator,
     ).images[0]
 
-    pipe.unet.set_attn_processor(AttnProcessor2_0())
-    base_image = pipe(
+
+    pipe.set_adapters(["res_adapter"], adapter_weights=[1.0])
+    res_adapt = pipe(
         prompt=prompt,
         negative_prompt=negative_prompt,
         prompt_2=prompt_2,
         negative_prompt_2=negative_prompt_2,
         width=width,
         height=height,
-        guidance_scale=guidance_scale_base,
-        num_inference_steps=num_inference_steps_base,
+        num_inference_steps=num_inference_steps,
+        guidance_scale=guidance_scale,
+        output_type="pil",
         generator=generator,
-        output_type="pil").images[0]
+    ).images[0]
 
-    
-    
-    
-    return [res_adapt,base_image]
+    return [res_adapt, base_image]
 
 
 examples = [
-    "A realistic photograph of an astronaut in a jungle, cold color palette, detailed, 8k",
-    "An astronaut riding a green horse",
-    "cinematic film still, photo of a girl, cyberpunk, neonpunk, headset, city at night, sony fe 12-24mm f/2.8 gm, close up, 32k uhd, wallpaper, analog film grain, SONY headset"
+    "A girl smiling",
+    "A boy smiling",
 ]
 
 theme = gr.themes.Base(
-    font=[gr.themes.GoogleFont('Libre Franklin'), gr.themes.GoogleFont('Public Sans'), 'system-ui', 'sans-serif'],
+    font=[
+        gr.themes.GoogleFont("Libre Franklin"),
+        gr.themes.GoogleFont("Public Sans"),
+        "system-ui",
+        "sans-serif",
+    ],
 )
 with gr.Blocks(css="footer{display:none !important}", theme=theme) as demo:
     gr.Markdown(DESCRIPTION)
@@ -136,13 +156,15 @@ with gr.Blocks(css="footer{display:none !important}", theme=theme) as demo:
     # result = gr.Gallery(label="Right is Res-Adapt-LORA and Left is Base"),
     with gr.Accordion("Advanced options", open=False):
         with gr.Row():
-            use_negative_prompt = gr.Checkbox(label="Use negative prompt", value=True)
+            use_negative_prompt = gr.Checkbox(label="Use negative prompt", value=False)
             use_prompt_2 = gr.Checkbox(label="Use prompt 2", value=False)
-            use_negative_prompt_2 = gr.Checkbox(label="Use negative prompt 2", value=False)
+            use_negative_prompt_2 = gr.Checkbox(
+                label="Use negative prompt 2", value=False
+            )
         negative_prompt = gr.Text(
             label="Negative prompt",
             max_lines=1,
-            placeholder="ugly, deformed, noisy, blurry, nsfw, low contrast, text, BadDream, 3d, cgi, render, fake, anime, open mouth, big forehead, long neck",
+            placeholder="Enter your prompt",
             visible=True,
         )
         prompt_2 = gr.Text(
@@ -182,19 +204,19 @@ with gr.Blocks(css="footer{display:none !important}", theme=theme) as demo:
                 value=512,
             )
         with gr.Row():
-            guidance_scale_base = gr.Slider(
-                label="Guidance scale for base",
-                minimum=1,
+            guidance_scale = gr.Slider(
+                label="Guidance scale",
+                minimum=0,
                 maximum=20,
                 step=0.1,
-                value=9.5,
+                value=0,
             )
-            num_inference_steps_base = gr.Slider(
-                label="Number of inference steps for base",
-                minimum=10,
-                maximum=100,
+            num_inference_steps = gr.Slider(
+                label="Number of inference steps",
+                minimum=1,
+                maximum=50,
                 step=1,
-                value=25,
+                value=4,
             )
     gr.Examples(
         examples=examples,
@@ -251,12 +273,12 @@ with gr.Blocks(css="footer{display:none !important}", theme=theme) as demo:
             seed,
             width,
             height,
-            guidance_scale_base,
-            num_inference_steps_base,
+            guidance_scale,
+            num_inference_steps,
         ],
-        outputs=gr.Gallery(label="Left is Res-Adapt-LORA and Right is Base"),
+        outputs=gr.Gallery(label="Left is ResAdapter and Right is Base"),
         api_name="run",
     )
 
 if __name__ == "__main__":
-    demo.queue(max_size=20, api_open=False).launch(show_api=False)
+    demo.queue(max_size=20, api_open=False).launch(show_api=False)