app.py

import torch
import spaces
import gradio as gr
import random
import numpy as np
from diffusers import ZImagePipeline

# Load the pipeline once at startup
print("Loading Z-Image-Turbo pipeline...")
pipe = ZImagePipeline.from_pretrained(
    "Tongyi-MAI/Z-Image-Turbo",
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=False,
)
pipe_no_lora = ZImagePipeline.from_pretrained(
    "Tongyi-MAI/Z-Image-Turbo",
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=False,
)
pipe.load_lora_weights("Shakker-Labs/AWPortrait-Z", weight_name="AWPortrait-Z.safetensors", adapter_name="lora")
pipe.set_adapters(["lora",], adapter_weights=[1.])
pipe.fuse_lora(adapter_names=["lora"], lora_scale=.9)
pipe.unload_lora_weights()
pipe.to("cuda")
pipe_no_lora.to("cuda")

# ======== AoTI compilation + FA3 ========
pipe.transformer.layers._repeated_blocks = ["ZImageTransformerBlock"]
spaces.aoti_blocks_load(pipe.transformer.layers, "zerogpu-aoti/Z-Image", variant="fa3")

pipe_no_lora.transformer.layers._repeated_blocks = ["ZImageTransformerBlock"]
spaces.aoti_blocks_load(pipe_no_lora.transformer.layers, "zerogpu-aoti/Z-Image", variant="fa3")
MAX_SEED = np.iinfo(np.int32).max
print("Pipeline loaded!")

@spaces.GPU
def generate_image(prompt, height, width, num_inference_steps, seed=42, randomize_seed=True, progress=gr.Progress(track_tqdm=True)):
    """Generate an image from the given prompt."""
    if randomize_seed:
        seed = random.randint(0, MAX_SEED)
    image = pipe(
        prompt=prompt,
        height=int(height),
        width=int(width),
        num_inference_steps=int(num_inference_steps),
        guidance_scale=0.0,  # Guidance should be 0 for Turbo models
        generator = torch.Generator(device="cuda").manual_seed(seed)
    ).images[0]

    image_no_lora = pipe_no_lora(
        prompt=prompt,
        height=int(height),
        width=int(width),
        num_inference_steps=int(num_inference_steps),
        guidance_scale=0.0,  # Guidance should be 0 for Turbo models
        generator = torch.Generator(device="cuda").manual_seed(seed)
    ).images[0]
    
    return (image_no_lora,image), seed


# Example prompts
examples = [
    ["A dramatic close-up high-fashion portrait with avant-garde futuristic styling, metallic accents, sculptural makeup, glowing rim light, hyperreal detail, cool-toned color palette, glossy finish, fashion campaign quality."],
]
css = """
#col-container { max-width: 950px; margin: 0 auto; }
.dark .progress-text { color: white !important; }

#examples { max-width: 950px; margin: 0 auto; }

.dark #examples button,
.dark #examples .example,
.dark #examples span {
    color: white !important;
}
"""
# Build the Gradio interface
with gr.Blocks() as demo:
    
    with gr.Column(elem_id="col-container"):
        gr.Markdown(
        """
        # Z-Image-Turbo Portrait✨
        
        Generate high-quality portrait images with [Tongyi-MAI/Z-Image-Turbo](https://huggingface.co/Tongyi-MAI/Z-Image-Turbo) using [portrait-beauty LoRA by @dynamicwangs and Shakker Labs](https://huggingface.co/Shakker-Labs/AWPortrait-Z), for fast inference with enhanced details. 
        This turbo model generates images in just 8 inference steps!
        """
    )
    
        with gr.Row():
            with gr.Column(scale=1):
                prompt = gr.Textbox(
                        label="Prompt",
                        placeholder="Enter your image description...",
                        max_lines=4,
                    )
                generate_btn = gr.Button("Generate", variant="primary")
                with gr.Accordion("Advanced Settings", open=False):
                    with gr.Row():
                        height = gr.Slider(
                            minimum=512,
                            maximum=2048,
                            value=1024,
                            step=64,
                            label="Height",
                        )
                        width = gr.Slider(
                            minimum=512,
                            maximum=2048,
                            value=1024,
                            step=64,
                            label="Width",
                        )
                    
                    with gr.Row():
                        num_inference_steps = gr.Slider(
                            minimum=1,
                            maximum=20,
                            value=9,
                            step=1,
                            label="Inference Steps",
                            info="9 steps results in 8 DiT forwards",
                        )
                    
                    with gr.Row():
                        seed = gr.Number(
                            label="Seed",
                            value=42,
                            precision=0,
                        )
                        randomize_seed = gr.Checkbox(
                            label="Randomize Seed",
                            value=True,
                        )
            with gr.Column(scale=1):    
                output_image = gr.ImageSlider(
                        label="Output (Left-with the LoRA, Right-without)",
                        type="pil",
                    )
            
                
        
        gr.Examples(
            examples=examples,
            inputs=[prompt],
            cache_examples=False,
            elem_id="examples", 
        )
    
        
    # Connect the generate button
    generate_btn.click(
        fn=generate_image,
        inputs=[prompt, height, width, num_inference_steps, seed, randomize_seed],
        outputs=[output_image, seed],
    )
    
    # Also allow generating by pressing Enter in the prompt box
    prompt.submit(
        fn=generate_image,
        inputs=[prompt, height, width, num_inference_steps, seed, randomize_seed],
        outputs=[output_image, seed],
    )

if __name__ == "__main__":
    demo.launch(mcp_server=True, theme=gr.themes.Citrus(), css=css)