Preloaded model and improved UI

app.py

@@ -1,3 +1,5 @@
+import os
+
 import gradio as gr
 import spaces
 import torch
@@ -5,7 +7,7 @@ from diffusers import AutoPipelineForImage2Image, StableDiffusionInstructPix2Pix
 from loguru import logger
 from PIL import Image
 
-models = [
+SUPPORTED_MODELS = [
     "stabilityai/sdxl-turbo",
     "stabilityai/stable-diffusion-3-medium-diffusers",
     "stabilityai/stable-diffusion-xl-refiner-1.0",
@@ -13,6 +15,12 @@ models = [
 ]
 DEFAULT_MODEL = "stabilityai/stable-diffusion-xl-refiner-1.0"
 
+MAX_IMAGE_SIZE = 1024
+
+
+model = os.environ.get("MODEL_ID", DEFAULT_MODEL)
+gpu_duration = int(os.environ.get("GPU_DURATION", 60))
+
 
 def load_pipeline(model):
     pipeline_type = (
@@ -22,96 +30,143 @@ def load_pipeline(model):
     )
 
     return pipeline_type.from_pretrained(
-        model,
-        torch_dtype=torch.float16,
-        use_safetensors=True,
-        variant="fp16"
+        model, torch_dtype=torch.float16, use_safetensors=True, variant="fp16"
     )
 
 
-load_pipeline(DEFAULT_MODEL).to("cuda")
-loaded_models = {DEFAULT_MODEL}
+pipe = load_pipeline(DEFAULT_MODEL).to("cuda")
 
 
-def generate_image(
-    model: str,
+@logger.catch(reraise=True)
+@spaces.GPU(duration=gpu_duration)
+def infer(
     prompt: str,
     init_image: Image.Image,
+    negative_prompt: str,
+    width: int,
+    height: int,
     strength: float,
-    progress,
+    num_inference_steps: int,
+    guidance_scale: float,
+    progress=gr.Progress(track_tqdm=True),
 ):
-    logger.debug(f"Loading pipeline: {dict(model=model)}")
-    pipe = load_pipeline(model).to("cuda")
+    logger.info(
+        f"Starting image generation: {dict(model=model, prompt=prompt, image=init_image)}"
+    )
+
+    # Downscale the image
+    init_image.thumbnail((1024, 1024))
 
     logger.debug(f"Generating image: {dict(prompt=prompt)}")
     additional_args = (
-        {} if model == "timbrooks/instruct-pix2pix" else dict(strength=strength)
+        {}
+        if model == "timbrooks/instruct-pix2pix" or strength == 0
+        else dict(strength=strength)
     )
 
-    def progress_callback(pipe, step_index, timestep, callback_kwargs):
-        logger.trace(
-            f"Callback: {dict(num_timesteps=pipe.num_timesteps, step_index=step_index, timestep=timestep)}"
-        )
-        progress((step_index + 1, pipe.num_timesteps))
-        return callback_kwargs
-
     images = pipe(
         prompt=prompt,
         image=init_image,
-        callback_on_step_end=progress_callback,
+        negative_prompt=negative_prompt,
+        width=width,
+        height=height,
+        num_inference_steps=num_inference_steps,
+        guidance_scale=guidance_scale,
         **additional_args,
     ).images
     return images[0]
 
 
-@spaces.GPU
-def gpu(*args, **kwargs):
-    return generate_image(*args, **kwargs)
-
-
-@spaces.GPU(duration=180)
-def gpu_3min(*args, **kwargs):
-    return generate_image(*args, **kwargs)
-
-
-@logger.catch(reraise=True)
-def generate(
-    model: str,
-    prompt: str,
-    init_image: Image.Image,
-    strength: float,
-    progress=gr.Progress(),
-):
-    logger.info(
-        f"Starting image generation: {dict(model=model, prompt=prompt, image=init_image, strength=strength)}"
+css = """
+#col-container {
+    margin: 0 auto;
+    max-width: 640px;
+}
+"""
+
+with gr.Blocks(css=css) as demo:
+    with gr.Column(elem_id="col-container"):
+        gr.Markdown("# Image-to-Image")
+        gr.Markdown(f"## Model: {model}")
+
+        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")
+
+        init_image = gr.Image(label="Initial image", type="pil")
+
+        result = gr.Image(label="Result", show_label=False)
+
+        with gr.Accordion("Advanced Settings", open=False):
+            negative_prompt = gr.Text(
+                label="Negative prompt",
+                max_lines=1,
+                placeholder="Enter a negative prompt",
+            )
+
+            with gr.Row():
+                width = gr.Slider(
+                    label="Width",
+                    minimum=256,
+                    maximum=MAX_IMAGE_SIZE,
+                    step=32,
+                    value=1024,
+                )
+
+                height = gr.Slider(
+                    label="Height",
+                    minimum=256,
+                    maximum=MAX_IMAGE_SIZE,
+                    step=32,
+                    value=1024,
+                )
+
+            with gr.Row():
+                strength = gr.Slider(
+                    label="Strength",
+                    minimum=0.0,
+                    maximum=1.0,
+                    step=0.01,
+                    value=0.0,
+                )
+
+                num_inference_steps = gr.Slider(
+                    label="Number of inference steps",
+                    minimum=1,
+                    maximum=100,
+                    step=1,
+                    value=50,
+                )
+
+                guidance_scale = gr.Slider(
+                    label="Guidance scale",
+                    minimum=0.0,
+                    maximum=100.0,
+                    step=0.1,
+                    value=0.0,
+                )
+    gr.on(
+        triggers=[run_button.click, prompt.submit],
+        fn=infer,
+        inputs=[
+            prompt,
+            init_image,
+            negative_prompt,
+            width,
+            height,
+            strength,
+            num_inference_steps,
+            guidance_scale,
+        ],
+        outputs=[result],
     )
 
-    # Downscale the image
-    init_image.thumbnail((1024, 1024))
-
-    # Cache the model files for the pipeline
-    if model not in loaded_models:
-        logger.debug(f"Caching pipeline: {dict(model=model)}")
-
-        load_pipeline(model)
-        loaded_models.add(model)
-
-    gpu_runner = gpu_3min if model == "timbrooks/instruct-pix2pix" else gpu
-
-    return gpu_runner(model, prompt, init_image, strength, progress)
-
-
-demo = gr.Interface(
-    fn=generate,
-    inputs=[
-        gr.Dropdown(
-            label="Model", choices=models, value=DEFAULT_MODEL, allow_custom_value=True
-        ),
-        gr.Text(label="Prompt"),
-        gr.Image(label="Init image", type="pil"),
-        gr.Slider(label="Strength", minimum=0, maximum=1, value=0.3),
-    ],
-    outputs=[gr.Image(label="Output")],
-)
-
-demo.launch()
+if __name__ == "__main__":
+    demo.launch()