app.py

import os
import torch
import spaces
import gradio as gr
from diffusers import DiffusionPipeline, EulerDiscreteScheduler

# ---------------------------------------------------------------------
# Model setup (maps roughly to UNETLoader + VAELoader + CLIPLoader)
# ---------------------------------------------------------------------

# Change this to your preferred SD3 model or a local path.
# For example, you can replace with a local snapshot inside the Space repo.
MODEL_ID = os.getenv("MODEL_ID", "Tongyi-MAI/Z-Image-Turbo")

device = "cuda" if torch.cuda.is_available() else "cpu"

pipe = DiffusionPipeline.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.bfloat16,
    use_safetensors=True,
)

# KSampler → choose a scheduler (Euler is close to your Comfy euler/simple)
pipe.to(device)
pipe.transformer.layers._repeated_blocks = ["ZImageTransformerBlock"]
spaces.aoti_blocks_load(pipe.transformer.layers, "zerogpu-aoti/Z-Image", variant="fa3")

# ---------------------------------------------------------------------
# Inference function (maps to CLIPTextEncode + EmptySD3LatentImage + KSampler + VAEDecode)
# ---------------------------------------------------------------------
@spaces.GPU
@spaces.GPU
def generate_images(
    positive: str,
    negative: str,
    width: int,
    height: int,
    steps: int,
    cfg: float,
    seed: int,
    num_images: int,
):
    run_device = "cuda" if torch.cuda.is_available() else "cpu"
    pipe.to(run_device)

    num_images = int(num_images)
    width = int(width)
    height = int(height)
    steps = int(steps)

    images = []

    # seed >= 0  -> deterministic series: seed, seed+1, ...
    # seed < 0   -> fully random seeds per image
    fixed_base_seed = int(seed) if seed >= 0 else None

    for i in range(num_images):
        if fixed_base_seed is None:
            # random seed for this image
            this_seed = torch.randint(0, 2**63 - 1, (1,), device=run_device).item()
        else:
            # deterministic offset
            this_seed = fixed_base_seed + i

        generator = torch.Generator(device=run_device).manual_seed(int(this_seed))

        out = pipe(
            prompt=positive,
            negative_prompt=negative or None,
            width=width,
            height=height,
            num_inference_steps=steps,
            guidance_scale=float(cfg),
            num_images_per_prompt=1,
            generator=generator,
        ).images[0]

        images.append(out)

    return images




# ---------------------------------------------------------------------
# Gradio UI (inputs correspond to Comfy node widgets_values)
# ---------------------------------------------------------------------
with gr.Blocks() as demo:
    gr.Markdown("# SD3 Text-to-Image – ComfyUI Workflow Port")

    with gr.Row():
        with gr.Column():
            positive = gr.Textbox(
                label="Positive Prompt",
                value="masterpiece, best quality, extremely detailed, high resolution.",  # from CLIP Text Encode (Positive Prompt)
                lines=5,
            )
            negative = gr.Textbox(
                label="Negative Prompt",
                value="watermark, blurry, ugly, bad anatomy",  # from CLIP Text Encode (Negative Prompt)
                lines=4,
            )

            width = gr.Slider(
                label="Width",
                minimum=256,
                maximum=1536,
                step=64,
                value=512,  # EmptySD3LatentImage width
            )
            height = gr.Slider(
                label="Height",
                minimum=256,
                maximum=1536,
                step=64,
                value=768,  # EmptySD3LatentImage height
            )

            steps = gr.Slider(
                label="Steps (KSampler)",
                minimum=1,
                maximum=50,
                step=1,
                value=12,  # KSampler steps
            )
            cfg = gr.Slider(
                label="CFG (Guidance Scale)",
                minimum=1.0,
                maximum=20.0,
                step=0.1,
                value=1.5,  # KSampler cfg in your graph
            )
            num_images = gr.Slider(
                label="Batch Size",
                minimum=1,
                maximum=8,
                step=1,
                value=6,  # EmptySD3LatentImage batch_size
            )
            seed = gr.Number(
                label="Seed (negative for random)",
                value=-1,  # "randomize" in Comfy
                precision=0,
            )

            run_btn = gr.Button("Generate")

        with gr.Column():
            gallery = gr.Gallery(
                label="Output Images",
                show_label=True,
                columns=3,
                height=768,
                object_fit="contain",  # keep full image visible in cell
                preview=False,         # do not start in zoomed preview mode
                allow_preview=True,    # still allow zoom when clicked
            )


    run_btn.click(
        fn=generate_images,
        inputs=[positive, negative, width, height, steps, cfg, seed, num_images],
        outputs=[gallery],
    )

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