app.py

import gradio as gr
import numpy as np
import random
import spaces
import torch
import math
# from diffusers import ZImagePipeline
from diffusers import QwenImagePipeline, FlowMatchEulerDiscreteScheduler


import os
import requests
import tempfile
import shutil
from urllib.parse import urlparse

dtype = torch.bfloat16
device = "cuda" if torch.cuda.is_available() else "cpu"




scheduler_config = {
    "base_image_seq_len": 256,
    "base_shift": math.log(3),  # We use shift=3 in distillation
    "invert_sigmas": False,
    "max_image_seq_len": 8192,
    "max_shift": math.log(3),  # We use shift=3 in distillation
    "num_train_timesteps": 1000,
    "shift": 1.0,
    "shift_terminal": None,  # set shift_terminal to None
    "stochastic_sampling": False,
    "time_shift_type": "exponential",
    "use_beta_sigmas": False,
    "use_dynamic_shifting": True,
    "use_exponential_sigmas": False,
    "use_karras_sigmas": False,
}
scheduler = FlowMatchEulerDiscreteScheduler.from_config(scheduler_config)

# Load the model pipeline
pipe = QwenImagePipeline.from_pretrained("Qwen/Qwen-Image-2512", 
                                         scheduler=scheduler, 
                                         torch_dtype=dtype).to(device)
pipe.load_lora_weights(
    "lightx2v/Qwen-Image-2512-Lightning", weight_name="Qwen-Image-2512-Lightning-4steps-V1.0-bf16.safetensors"
)
# pipe.load_lora_weights("Wuli-art/Qwen-Image-2512-Turbo-LoRA")
pipe.fuse_lora()

pipe.vae.enable_tiling() 

# Load the model pipeline
# pipe = QwenImagePipeline.from_pretrained("Qwen/Qwen-Image-2512", torch_dtype=dtype).to(device)
# pipe.vae.enable_tiling() 
# pipe.load_lora_weights("Wuli-art/Qwen-Image-2512-Turbo-LoRA")
# pipe.fuse_lora()

torch.cuda.empty_cache()

MAX_SEED = np.iinfo(np.int32).max
MAX_IMAGE_SIZE = 2048

# pipe.flux_pipe_call_that_returns_an_iterable_of_images = flux_pipe_call_that_returns_an_iterable_of_images.__get__(pipe)

# flymy-ai/qwen-image-realism-lora

def load_lora_auto(pipe, lora_input):
    lora_input = lora_input.strip()
    if not lora_input:
        return

    # If it's just an ID like "author/model"
    if "/" in lora_input and not lora_input.startswith("http"):
        pipe.load_lora_weights(lora_input)
        return

    if lora_input.startswith("http"):
        url = lora_input

        # Repo page (no blob/resolve)
        if "huggingface.co" in url and "/blob/" not in url and "/resolve/" not in url:
            repo_id = urlparse(url).path.strip("/")
            pipe.load_lora_weights(repo_id)
            return

        # Blob link → convert to resolve link
        if "/blob/" in url:
            url = url.replace("/blob/", "/resolve/")

        # Download direct file
        tmp_dir = tempfile.mkdtemp()
        local_path = os.path.join(tmp_dir, os.path.basename(urlparse(url).path))

        try:
            print(f"Downloading LoRA from {url}...")
            resp = requests.get(url, stream=True)
            resp.raise_for_status()
            with open(local_path, "wb") as f:
                for chunk in resp.iter_content(chunk_size=8192):
                    f.write(chunk)
            print(f"Saved LoRA to {local_path}")
            pipe.load_lora_weights(local_path)
        finally:
            shutil.rmtree(tmp_dir, ignore_errors=True)

@spaces.GPU()
def infer(prompt, negative_prompt, seed=42, randomize_seed=False, width=1024, height=1024, guidance_scale=4, num_inference_steps=28, lora_id=None, lora_scale=0.95, progress=gr.Progress(track_tqdm=True)):
    if randomize_seed:
        seed = random.randint(0, MAX_SEED)
    generator = torch.Generator().manual_seed(seed)

    
    if lora_id and lora_id.strip() != "":
        pipe.unload_lora_weights()
        load_lora_auto(pipe, lora_id)
    
    try:
        image = pipe(
        prompt=prompt,
        negative_prompt=negative_prompt,
        width=width,
        height=height,
        num_inference_steps=num_inference_steps,
        generator=generator,
        true_cfg_scale=guidance_scale,  # Use a fixed default for distilled guidance
        guidance_scale=1.0
    ).images[0]
        print("Image Generation Completed for: ", prompt, lora_id)
        return image, seed
    finally:
        # Unload LoRA weights if they were loaded
        if lora_id:
            pipe.unload_lora_weights()


lora_examples = [
    ["prithivMLmods/Qwen-Image-2512-Pixel-Art-LoRA"],
    ["RayyanAhmed9477/qwen-image-2512-lora-advertisement"],
    ["lichorosario/t21person-qwen-image-2512"],
    ["https://huggingface.co/lichorosario/flat-cartoon-qwen-image-2512/resolve/main/flat-cartoon-qwen-image-2512_20.safetensors"]
]
    
examples = [
    "a tiny astronaut hatching from an egg on the moon",
    "a cat holding a sign that says hello world",
    "an anime illustration of a wiener schnitzel",
]
    
css = """
#col-container {
   margin: 0 auto;
   max-width: 960px;
}
.generate-btn {
   background: linear-gradient(90deg, #4B79A1 0%, #283E51 100%) !important;
   border: none !important;
   color: white !important;
}
.generate-btn:hover {
   transform: translateY(-2px);
   box-shadow: 0 5px 15px rgba(0,0,0,0.2);
}
"""

with gr.Blocks(css=css, delete_cache=(3600, 7200)) as app:  # cleanup: check every 1h, delete files >2h old
    gr.HTML("<center><h1>Qwen-Image 2512 with LoRA support</h1></center>")
    with gr.Column(elem_id="col-container"):
        with gr.Row():
            with gr.Column():
                with gr.Row():
                    text_prompt = gr.Textbox(label="Prompt", placeholder="Enter a prompt here", lines=3, elem_id="prompt-text-input")
                with gr.Row():
                    custom_lora = gr.Textbox(label="Custom LoRA (optional)", info="URL or the path to the LoRA weights", placeholder="ostris/z_image_turbo_childrens_drawings")

                with gr.Row():
                    with gr.Accordion("Advanced Settings", open=False):
                        lora_scale = gr.Slider(
                            label="LoRA Scale",
                            minimum=0,
                            maximum=2,
                            step=0.01,
                            value=1,
                        )
                        negative_prompt = gr.Textbox(label="Negative prompt", value = "低分辨率，低画质，肢体畸形，手指畸形，画面过饱和，蜡像感，人脸无细节，过度光滑，画面具有AI感。构图混乱。文字模糊，扭曲", lines=3, elem_id="prompt-text-input")
                        with gr.Row():
                            width = gr.Slider(label="Width", value=1328, minimum=64, maximum=2048, step=16)
                            height = gr.Slider(label="Height", value=1328, minimum=64, maximum=2048, step=16)
                        seed = gr.Slider(label="Seed", value=-1, minimum=-1, maximum=4294967296, step=1)
                        randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
                        with gr.Row():
                            steps = gr.Slider(label="Inference steps steps", value=4, minimum=1, maximum=20, step=1)
                            cfg = gr.Slider(label="Guidance Scale", value=1, minimum=0, maximum=20, step=0.5)
                        # method = gr.Radio(label="Sampling method", value="DPM++ 2M Karras", choices=["DPM++ 2M Karras", "DPM++ SDE Karras", "Euler", "Euler a", "Heun", "DDIM"])
                

                with gr.Row():
                    # text_button = gr.Button("Run", variant='primary', elem_id="gen-button")
                    text_button = gr.Button("✨ Generate Image", variant='primary', elem_classes=["generate-btn"])

                
            with gr.Column():
                with gr.Row():
                    image_output = gr.Image(type="pil", label="Image Output", elem_id="gallery")
                gr.Examples(
                    examples=lora_examples,
                    inputs=[custom_lora],
                    label="Select a LoRA Model",
                )
        
        # gr.Markdown(article_text)
        with gr.Column():
            gr.Examples(
                examples = examples,
                inputs = [text_prompt],
            )
    gr.on(
        triggers=[text_button.click, text_prompt.submit],
        fn = infer,
        inputs=[text_prompt, negative_prompt, seed, randomize_seed, width, height, cfg, steps, custom_lora, lora_scale], 
        outputs=[image_output, seed]
    )
        
        # text_button.click(query, inputs=[custom_lora, text_prompt, steps, cfg, randomize_seed, seed, width, height], outputs=[image_output,seed_output, seed])
        # text_button.click(infer, inputs=[text_prompt, seed, randomize_seed, width, height, cfg, steps, custom_lora, lora_scale], outputs=[image_output,seed_output, seed])

app.launch(mcp_server=True, share=True)