app.py

import gradio as gr
import torch
import spaces
from diffusers import DiffusionPipeline
from PIL import Image
from typing import List, Optional, Any

# --- Model Configuration ---
MODEL_V1 = "CompVis/stable-diffusion-v1-4"
MODEL_V2 = "Manojb/stable-diffusion-2-1-base"
DEVICE = "cuda"

# Use bfloat16 for optimized performance on modern GPUs (H200/A100/H100)
DTYPE = torch.bfloat16

# Default prompts for generation when user input is empty
DEFAULT_PROMPT_V1 = "A stunning photorealistic image of a golden retriever wearing a crown, in a grand hall, cinematic lighting, masterpiece, 4k"
DEFAULT_PROMPT_V2 = "A detailed matte painting of an ancient ruined city overgrown with vines, dramatic sunset, fantasy art, 8k, cinematic"

print("Loading Models...")
pipe_v1 = DiffusionPipeline.from_pretrained(
    MODEL_V1,
    torch_dtype=DTYPE,
    safety_checker=None,
    requires_safety_checker=False,
    # Use from_single_file=True if loading .ckpt or .safetensors files directly
).to(DEVICE)

pipe_v2 = DiffusionPipeline.from_pretrained(
    MODEL_V2,
    torch_dtype=DTYPE,
    safety_checker=None,
    requires_safety_checker=False,
).to(DEVICE)
print("Models Loaded.")


@spaces.GPU(duration=1500)
def compile_optimized_models():
    """
    Performs Ahead-of-Time (AoT) compilation for improved ZeroGPU performance.
    """
    # --- Compilation for SD 1.4 (pipe_v1) ---
    print(f"Compiling UNet for {MODEL_V1} (SD 1.4)...")
    try:
        with spaces.aoti_capture(pipe_v1.unet) as call:
            # Run a quick example call (512x512, low steps) to capture inputs
            pipe_v1(
                prompt="compilation test",
                num_inference_steps=2,
                height=512, width=512
            )
        exported_v1 = torch.export.export(pipe_v1.unet, args=call.args, kwargs=call.kwargs)
        compiled_v1 = spaces.aoti_compile(exported_v1)
        spaces.aoti_apply(compiled_v1, pipe_v1.unet)
        print(f"Compilation for {MODEL_V1} complete.")
    except Exception as e:
        print(f"Warning: AoT compilation failed for SD 1.4. Running unoptimized. Error: {e}")

    # --- Compilation for SD 2.1 Base (pipe_v2) ---
    print(f"Compiling UNet for {MODEL_V2} (SD 2.1 Base)...")
    try:
        with spaces.aoti_capture(pipe_v2.unet) as call:
            # Run a quick example call (512x512, low steps) to capture inputs
            pipe_v2(
                prompt="compilation test",
                num_inference_steps=2,
                height=512, width=512
            )
        exported_v2 = torch.export.export(pipe_v2.unet, args=call.args, kwargs=call.kwargs)
        compiled_v2 = spaces.aoti_compile(exported_v2)
        spaces.aoti_apply(compiled_v2, pipe_v2.unet)
        print(f"Compilation for {MODEL_V2} complete.")
    except Exception as e:
        print(f"Warning: AoT compilation failed for SD 2.1 Base. Running unoptimized. Error: {e}")

# Run compilation once at startup
compile_optimized_models()


@spaces.GPU
def generate(
    model_choice: str,
    prompt: str,
    guidance_scale: float,
    num_inference_steps: int
) -> List[Image.Image]:
    """Generates images using the selected Stable Diffusion model."""

    if model_choice == MODEL_V1:
        pipe = pipe_v1
        if not prompt:
            prompt = DEFAULT_PROMPT_V1
    elif model_choice == MODEL_V2:
        pipe = pipe_v2
        if not prompt:
            prompt = DEFAULT_PROMPT_V2
    else:
        raise gr.Error("Invalid model selection.")

    # We must use the resolution used during AoT compilation (512x512)
    # for best performance.
    result = pipe(
        prompt=prompt,
        guidance_scale=guidance_scale,
        num_inference_steps=num_inference_steps,
        num_images_per_prompt=4,  # Generate 4 images as implied by gallery output
        height=512,
        width=512
    ).images

    return result


def display_uploads(files: Optional[List[Any]]) -> List[str]:
    """Converts uploaded FileData objects to displayable paths."""
    if files:
        # FileData objects have a .path attribute pointing to the temporary file location
        return [f.path for f in files]
    return []


# --- Gradio Interface ---
with gr.Blocks(title="Stable Diffusion Models Demo") as demo:
    gr.HTML(
        """
        <div style='text-align: center; max-width: 800px; margin: 0 auto;'>
            <h1>Stable Diffusion v1.4 vs 2.1 Base</h1>
            <p>Select a model and enter a prompt to generate up to 4 images. Empty prompts use a powerful default prompt.</p>
            <p><a href="https://huggingface.co/spaces/akhaliq/anycoder" target="_blank">Built with anycoder</a></p>
        </div>
        """
    )

    with gr.Row():
        with gr.Column(scale=1):
            model_choice = gr.Radio(
                choices=[MODEL_V1, MODEL_V2],
                value=MODEL_V2,
                label="Model Selection",
                info="Select the base Stable Diffusion version to use."
            )
            prompt = gr.Textbox(
                label="Prompt",
                placeholder="Enter your prompt here (or leave empty for default demo prompt)"
            )

            with gr.Accordion("Generation Parameters", open=True):
                guidance_scale = gr.Slider(
                    minimum=1.0, maximum=15.0, value=7.5, step=0.5, label="Guidance Scale",
                    info="Higher values push the generation closer to the prompt."
                )
                num_inference_steps = gr.Slider(
                    minimum=10, maximum=100, value=50, step=5, label="Inference Steps",
                    info="More steps lead to higher quality, but slower generation."
                )

            run_btn = gr.Button("Generate 4 Images", variant="primary")

            # Handling image uploads (for auxiliary display/reference)
            uploaded_files = gr.File(
                label="Upload Reference Images (Max 4)",
                file_count="multiple",
                file_types=['image'],
                max_files=4,
                interactive=True
            )
            upload_display = gr.Gallery(
                label="Uploaded Images for Reference",
                columns=4,
                object_fit="contain",
                height=150,
                allow_preview=False
            )
            uploaded_files.change(display_uploads, uploaded_files, upload_display)

        with gr.Column(scale=3):
            output_gallery = gr.Gallery(
                label="Generated Images (512x512)",
                columns=2,
                object_fit="contain",
                height=512,
                preview=True
            )

    run_btn.click(
        fn=generate,
        inputs=[
            model_choice,
            prompt,
            guidance_scale,
            num_inference_steps
        ],
        outputs=output_gallery
    )

if __name__ == "__main__":
    demo.launch()