Update app.py

app.py

@@ -1,120 +1,101 @@
 import gradio as gr
 import torch
-import spaces
-from diffusers import StableDiffusionXLPipeline, UNet2DConditionModel, EulerDiscreteScheduler
-from diffusers.pipelines.stable_diffusion import StableDiffusionSafetyChecker
-from diffusers.image_processor import VaeImageProcessor
-from transformers import CLIPImageProcessor
+from diffusers import StableDiffusionXLPipeline, EulerDiscreteScheduler
 from huggingface_hub import hf_hub_download
 from safetensors.torch import load_file
+import spaces
+import os
+from PIL import Image
 
-device = "cuda"
-dtype = torch.float16
+SAFETY_CHECKER = os.environ.get("SAFETY_CHECKER", "0") == "1"
 
+# Constants
 base = "stabilityai/stable-diffusion-xl-base-1.0"
 repo = "ByteDance/SDXL-Lightning"
-opts = {
-    "1 Step"  : ("sdxl_lightning_1step_unet_x0.safetensors", 1),
-    "2 Steps" : ("sdxl_lightning_2step_unet.safetensors", 2),
-    "4 Steps" : ("sdxl_lightning_4step_unet.safetensors", 4),
-    "8 Steps" : ("sdxl_lightning_8step_unet.safetensors", 8),
+checkpoints = {
+    "1-Step" : ["sdxl_lightning_1step_unet_x0.safetensors", 1],
+    "2-Step" : ["sdxl_lightning_2step_unet.safetensors", 2],
+    "4-Step" : ["sdxl_lightning_4step_unet.safetensors", 4],
+    "8-Step" : ["sdxl_lightning_8step_unet.safetensors", 8],
 }
 
-# Inference function.
-@spaces.GPU()
-def generate(prompt, option, progress=gr.Progress()):
-    print(prompt, option)
-    ckpt, step = opts[option]
-
-    progress(0, desc="Initializing the model")
 
-    # Main pipeline.
-    unet = UNet2DConditionModel.from_config(base, subfolder="unet")
-    pipe = StableDiffusionXLPipeline.from_pretrained(base, unet=unet, torch_dtype=dtype, variant="fp16").to(device, dtype)
-    pipe.unet.load_state_dict(load_file(hf_hub_download(repo, ckpt), device=device))
-    pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing", prediction_type="sample" if step == 1 else "epsilon")
+# Ensure model and scheduler are initialized in GPU-enabled function
+if torch.cuda.is_available():
+    pipe = StableDiffusionXLPipeline.from_pretrained(base, torch_dtype=torch.float16, variant="fp16").to("cuda")
 
-    # Safety checker.
-    safety_checker=StableDiffusionSafetyChecker.from_pretrained("CompVis/stable-diffusion-safety-checker").to(device, dtype)
-    feature_extractor=CLIPImageProcessor.from_pretrained("openai/clip-vit-base-patch32")
-    image_processor = VaeImageProcessor(vae_scale_factor=8)
+if SAFETY_CHECKER:
+    from safety_checker import StableDiffusionSafetyChecker
+    from transformers import CLIPFeatureExtractor
 
-    def inference_callback(p, i, t, kwargs):
-        progress((i+1, step))
-        return kwargs
-    
-    # Inference loop.
-    progress((0, step))
-    results = pipe(prompt, num_inference_steps=step, guidance_scale=0, callback_on_step_end=inference_callback, output_type="pt")
-
-    # Safety check.
-    feature_extractor_input = image_processor.postprocess(results.images, output_type="pil")
-    safety_checker_input = feature_extractor(feature_extractor_input, return_tensors="pt")
-    pixel_values = safety_checker_input.pixel_values.to(device, dtype)
-    images, has_nsfw_concept = safety_checker(
-        images=results.images, clip_input=pixel_values
+    safety_checker = StableDiffusionSafetyChecker.from_pretrained(
+        "CompVis/stable-diffusion-safety-checker"
+    ).to("cuda")
+    feature_extractor = CLIPFeatureExtractor.from_pretrained(
+        "openai/clip-vit-base-patch32"
     )
-    if has_nsfw_concept[0]:
-        print(f"Safety checker triggered on prompt: {prompt}")
-    return images[0]
 
-with gr.Blocks(css="style.css") as demo:
-    gr.HTML(
-        "<h1><center>SDXL-Lightning</center></h1>" +
-        "<p><center>Lightning-fast text-to-image generation</center></p>" +
-        "<p><center><a href='https://huggingface.co/ByteDance/SDXL-Lightning'>https://huggingface.co/ByteDance/SDXL-Lightning</a></center></p>"
-    )
-    
-    with gr.Row():
-        prompt = gr.Textbox(
-            label="Text prompt",
-            scale=8
-        )
-        option = gr.Dropdown(
-            label="Inference steps",
-            choices=["1 Step", "2 Steps", "4 Steps", "8 Steps"],
-            value="4 Steps",
-            interactive=True
-        )
-        submit = gr.Button(
-            scale=1,
-            variant="primary"
+    def check_nsfw_images(
+        images: list[Image.Image],
+    ) -> tuple[list[Image.Image], list[bool]]:
+        safety_checker_input = feature_extractor(images, return_tensors="pt").to("cuda")
+        has_nsfw_concepts = safety_checker(
+            images=[images],
+            clip_input=safety_checker_input.pixel_values.to("cuda")
         )
-    
-    img = gr.Image(label="SDXL-Lightning Generated Image")
 
-    prompt.submit(
-        fn=generate,
-        inputs=[prompt, option],
-        outputs=img,
-    )
-    submit.click(
-        fn=generate,
-        inputs=[prompt, option],
-        outputs=img,
-    )
+        return images, has_nsfw_concepts
 
-    gr.Examples(
-        fn=generate,
-        examples=[
-            ["An owl perches quietly on a twisted branch deep within an ancient forest.", "1 Step"],
-            ["A lion in the galaxy, octane render", "2 Steps"],
-            ["A dolphin leaps through the waves, set against a backdrop of bright blues and teal hues.", "2 Steps"],
-            ["A girl smiling", "4 Steps"],
-            ["An astronaut riding a horse", "4 Steps"],
-            ["A fish on a bicycle, colorful art", "4 Steps"],
-            ["A close-up of an Asian lady with sunglasses.", "4 Steps"],
-            ["Rabbit portrait in a forest, fantasy", "4 Steps"],
-            ["A panda swimming", "4 Steps"],
-            ["Man portrait, ethereal", "8 Steps"],
-        ],
-        inputs=[prompt, option],
-        outputs=img,
-        cache_examples=False,
-    )
+# Function 
+@spaces.GPU(enable_queue=True)
+def generate_image(prompt, ckpt):
 
-    gr.HTML(
-        "<p><small><center>This demo is built together by the community</center></small></p>"
-    )
+    checkpoint = checkpoints[ckpt][0]
+    num_inference_steps = checkpoints[ckpt][1] 
+
+    if num_inference_steps==1:
+        # Ensure sampler uses "trailing" timesteps and "sample" prediction type for 1-step inference.
+        pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing", prediction_type="sample")
+    else:
+        # Ensure sampler uses "trailing" timesteps.
+        pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing")
+        
+    pipe.unet.load_state_dict(load_file(hf_hub_download(repo, checkpoint), device="cuda"))
+    results = pipe(prompt, num_inference_steps=num_inference_steps, guidance_scale=0)
+
+    if SAFETY_CHECKER:
+        images, has_nsfw_concepts = check_nsfw_images(results.images)
+        if any(has_nsfw_concepts):
+            gr.Warning("NSFW content detected.")
+            return Image.new("RGB", (512, 512))
+        return images[0]
+    return results.images[0]
+
+
+
+# Gradio Interface
+description = """
+This demo utilizes the SDXL-Lightning model by ByteDance, which is a lightning-fast text-to-image generative model capable of producing high-quality images in 4 steps.
+As a community effort, this demo was put together by AngryPenguin. Link to model: https://huggingface.co/ByteDance/SDXL-Lightning
+"""
+
+with gr.Blocks(css="style.css") as demo:
+    gr.HTML("<h1><center>Text-to-Image with SDXL-Lightning ⚡</center></h1>")
+    gr.Markdown(description)
+    with gr.Group():
+        with gr.Row():
+            prompt = gr.Textbox(label='Enter you image prompt:', scale=8)
+            ckpt = gr.Dropdown(label='Select inference steps',choices=['1-Step', '2-Step', '4-Step', '8-Step'], value='4-Step', interactive=True)
+            submit = gr.Button(scale=1, variant='primary')
+    img = gr.Image(label='SDXL-Lightning Generated Image')
+
+    prompt.submit(fn=generate_image,
+                 inputs=[prompt, ckpt],
+                 outputs=img,
+                 )
+    submit.click(fn=generate_image,
+                 inputs=[prompt, ckpt],
+                 outputs=img,
+                 )
     
 demo.queue().launch()