app.py

import gradio as gr
import numpy as np
import random
import torch
import spaces

from PIL import Image
from diffusers import FlowMatchEulerDiscreteScheduler
from optimization import optimize_pipeline_
from qwenimage.pipeline_qwenimage_edit_plus import QwenImageEditPlusPipeline
from qwenimage.transformer_qwenimage import QwenImageTransformer2DModel
from qwenimage.qwen_fa3_processor import QwenDoubleStreamAttnProcessorFA3

import math
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file

from PIL import Image
import os
import gradio as gr
from gradio_client import Client, handle_file
import tempfile
from huggingface_hub import InferenceClient


# --- Model Loading ---
dtype = torch.bfloat16
device = "cuda" if torch.cuda.is_available() else "cpu"

pipe = QwenImageEditPlusPipeline.from_pretrained("Qwen/Qwen-Image-Edit-2509", 
                                                transformer= QwenImageTransformer2DModel.from_pretrained("linoyts/Qwen-Image-Edit-Rapid-AIO", 
                                                                                                         subfolder='transformer',
                                                                                                         torch_dtype=dtype,
                                                                                                         device_map='cuda'),torch_dtype=dtype).to(device)

# Load the relight LoRA
pipe.load_lora_weights(
        "dx8152/Qwen-Image-Edit-2509-Relight", 
        weight_name="Qwen-Edit-Relight.safetensors", adapter_name="relight"
    )

pipe.set_adapters(["relight"], adapter_weights=[1.])
pipe.fuse_lora(adapter_names=["relight"], lora_scale=1.25)
pipe.unload_lora_weights()

pipe.transformer.__class__ = QwenImageTransformer2DModel
pipe.transformer.set_attn_processor(QwenDoubleStreamAttnProcessorFA3())

optimize_pipeline_(pipe, image=[Image.new("RGB", (1024, 1024)), Image.new("RGB", (1024, 1024))], prompt="prompt")


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

# Initialize translation client
translation_client = InferenceClient(
    api_key=os.environ.get("HF_TOKEN"),
)

def translate_to_chinese(text: str) -> str:
    """Translate any language text to Chinese using Qwen API."""
    if not text or not text.strip():
        return ""
    
    # Check if text is already primarily Chinese
    chinese_chars = sum(1 for char in text if '\u4e00' <= char <= '\u9fff')
    if chinese_chars / max(len(text), 1) > 0.5:
        # Already mostly Chinese, return as is
        return text
    
    try:
        completion = translation_client.chat.completions.create(
            model="Qwen/Qwen3-Next-80B-A3B-Instruct:novita",
            messages=[
                {
                    "role": "system",
                    "content": "You are a professional translator. Translate the user's text to Chinese. Only output the translated text, nothing else."
                },
                {
                    "role": "user",
                    "content": f"Translate this to Chinese: {text}"
                }
            ],
            max_tokens=500,
        )
        
        translated = completion.choices[0].message.content.strip()
        print(f"Translated '{text}' to '{translated}'")
        return translated
    except Exception as e:
        print(f"Translation error: {e}")
        # Fallback to original text if translation fails
        return text

def _generate_video_segment(input_image_path: str, output_image_path: str, prompt: str, request: gr.Request) -> str:
    """Generates a single video segment using the external service."""
    x_ip_token = request.headers['x-ip-token']
    video_client = Client("multimodalart/wan-2-2-first-last-frame", headers={"x-ip-token": x_ip_token})
    result = video_client.predict(
        start_image_pil=handle_file(input_image_path),
        end_image_pil=handle_file(output_image_path),
        prompt=prompt, api_name="/generate_video",
    )
    return result[0]["video"]

def build_relight_prompt(light_type, light_direction, light_intensity, prompt):
    """Build the relighting prompt based on user selections."""
    
    # Priority 1: User's prompt (translated to Chinese if needed)
    if prompt and prompt.strip():
        translated = translate_to_chinese(prompt)
        # Add trigger word if not already present
        if "重新照明" not in translated:
            return f"重新照明,{translated}"
        return translated
    
    # Priority 2: Build from controls
    prompt_parts = ["重新照明"]
    
    # Light type descriptions
    light_descriptions = {
        "soft_window": "使用窗帘透光（柔和漫射）的光线",  # Soft diffuse light from curtains
        "golden_hour": "使用金色黄昏的温暖光线",  # Warm golden hour light
        "studio": "使用专业摄影棚的均匀光线",  # Professional studio lighting
        "dramatic": "使用戏剧性的高对比度光线",  # Dramatic high-contrast lighting
        "natural": "使用自然日光",  # Natural daylight
        "neon": "使用霓虹灯光效果",  # Neon lighting effect
        "candlelight": "使用烛光的温暖氛围",  # Warm candlelight ambiance
        "moonlight": "使用月光的冷色调",  # Cool-toned moonlight
    }
    
    # Direction descriptions
    direction_descriptions = {
        "front": "从正面照射",  # From the front
        "side": "从侧面照射",  # From the side
        "back": "从背后照射",  # From behind (backlight)
        "top": "从上方照射",  # From above
        "bottom": "从下方照射",  # From below
    }
    
    # Intensity descriptions
    intensity_descriptions = {
        "soft": "柔和强度",  # Soft intensity
        "medium": "中等强度",  # Medium intensity
        "strong": "强烈强度",  # Strong intensity
    }
    
    # Build the prompt
    if light_type != "none":
        prompt_parts.append(light_descriptions.get(light_type, ""))
    
    if light_direction != "none":
        prompt_parts.append(direction_descriptions.get(light_direction, ""))
    
    if light_intensity != "none":
        prompt_parts.append(intensity_descriptions.get(light_intensity, ""))
    
    final_prompt = ",".join([p for p in prompt_parts if p])
    
    # Add instruction if we have settings
    if len(prompt_parts) > 1:
        final_prompt += "对图片进行重新照明"  # Relight the image
    
    return final_prompt if len(prompt_parts) > 1 else "重新照明,使用自然光线对图片进行重新照明"


@spaces.GPU
def infer_relight(
    image,
    light_type,
    light_direction,
    light_intensity,
    prompt,
    seed,
    randomize_seed,
    true_guidance_scale,
    num_inference_steps,
    height,
    width,
    prev_output = None,
    progress=gr.Progress(track_tqdm=True)
):
    final_prompt = build_relight_prompt(light_type, light_direction, light_intensity, prompt)
    print(f"Generated Prompt: {final_prompt}")

    if randomize_seed:
        seed = random.randint(0, MAX_SEED)
    generator = torch.Generator(device=device).manual_seed(seed)

    # Choose input image (prefer uploaded, else last output)
    pil_images = []
    if image is not None:
        if isinstance(image, Image.Image):
            pil_images.append(image.convert("RGB"))
        elif hasattr(image, "name"):
            pil_images.append(Image.open(image.name).convert("RGB"))
    elif prev_output:
        pil_images.append(prev_output.convert("RGB"))

    if len(pil_images) == 0:
        raise gr.Error("Please upload an image first.")

    result = pipe(
        image=pil_images,
        prompt=final_prompt,
        height=height if height != 0 else None,
        width=width if width != 0 else None,
        num_inference_steps=num_inference_steps,
        generator=generator,
        true_cfg_scale=true_guidance_scale,
        num_images_per_prompt=1,
    ).images[0]

    return result, seed, final_prompt

def create_video_between_images(input_image, output_image, prompt: str, request: gr.Request) -> str:
    """Create a video between the input and output images."""
    if input_image is None or output_image is None:
        raise gr.Error("Both input and output images are required to create a video.")
    
    try:
        
        with tempfile.NamedTemporaryFile(delete=False, suffix=".png") as tmp:
            input_image.save(tmp.name)
            input_image_path = tmp.name
        
        output_pil = Image.fromarray(output_image.astype('uint8'))
        with tempfile.NamedTemporaryFile(delete=False, suffix=".png") as tmp:
            output_pil.save(tmp.name)
            output_image_path = tmp.name
            
        video_path = _generate_video_segment(
            input_image_path, 
            output_image_path, 
            prompt if prompt else "Relighting transformation",
            request
        )
        return video_path
    except Exception as e:
        raise gr.Error(f"Video generation failed: {e}")


# --- UI ---
css = '''#col-container { max-width: 800px; margin: 0 auto; }
.dark .progress-text{color: white !important}
#examples{max-width: 800px; margin: 0 auto; }'''

def reset_all():
    return ["none", "none", "none", "", False, True]

def end_reset():
    return False

def update_dimensions_on_upload(image):
    if image is None:
        return 1024, 1024
    
    original_width, original_height = image.size
    
    if original_width > original_height:
        new_width = 1024
        aspect_ratio = original_height / original_width
        new_height = int(new_width * aspect_ratio)
    else:
        new_height = 1024
        aspect_ratio = original_width / original_height
        new_width = int(new_height * aspect_ratio)
        
    # Ensure dimensions are multiples of 8
    new_width = (new_width // 8) * 8
    new_height = (new_height // 8) * 8
    
    return new_width, new_height


with gr.Blocks(theme=gr.themes.Citrus(), css=css) as demo:
    with gr.Column(elem_id="col-container"):
        gr.Markdown("## 💡 Qwen Image Edit — Relighting Control")
        gr.Markdown("""
            Qwen Image Edit 2509 for Image Relighting ✨ 
            Using [dx8152's Qwen-Image-Edit-2509-Relight LoRA](https://huggingface.co/dx8152/Qwen-Image-Edit-2509-Relight) and [linoyts/Qwen-Image-Edit-Rapid-AIO](https://huggingface.co/linoyts/Qwen-Image-Edit-Rapid-AIO) for 4-step inference 💨
            """
        )

        with gr.Row():
            with gr.Column():
                image = gr.Image(label="Input Image", type="pil")
                prev_output = gr.Image(value=None, visible=False)
                is_reset = gr.Checkbox(value=False, visible=False)

                with gr.Tab("Lighting Controls"):
                    light_type = gr.Dropdown(
                        label="Light Type",
                        choices=[
                            ("None", "none"),
                            ("Soft Window Light (柔和窗光)", "soft_window"),
                            ("Golden Hour (金色黄昏)", "golden_hour"),
                            ("Studio Lighting (摄影棚灯光)", "studio"),
                            ("Dramatic (戏剧性)", "dramatic"),
                            ("Natural Daylight (自然日光)", "natural"),
                            ("Neon (霓虹灯)", "neon"),
                            ("Candlelight (烛光)", "candlelight"),
                            ("Moonlight (月光)", "moonlight"),
                        ],
                        value="none"
                    )
                    
                    light_direction = gr.Dropdown(
                        label="Light Direction",
                        choices=[
                            ("None", "none"),
                            ("Front (正面)", "front"),
                            ("Side (侧面)", "side"),
                            ("Back (背光)", "back"),
                            ("Top (上方)", "top"),
                            ("Bottom (下方)", "bottom"),
                        ],
                        value="none"
                    )
                    
                    light_intensity = gr.Dropdown(
                        label="Light Intensity",
                        choices=[
                            ("None", "none"),
                            ("Soft (柔和)", "soft"),
                            ("Medium (中等)", "medium"),
                            ("Strong (强烈)", "strong"),
                        ],
                        value="none"
                    )
                
                with gr.Tab("Custom Prompt"):
                    prompt = gr.Textbox(
                        label="Relighting Prompt",
                        placeholder="Example: Add warm sunset lighting from the right",
                        lines=3
                    )
                
                with gr.Row():
                    reset_btn = gr.Button("Reset")
                    run_btn = gr.Button("Generate", variant="primary")

                with gr.Accordion("Advanced Settings", open=False):
                    seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0)
                    randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)
                    true_guidance_scale = gr.Slider(label="True Guidance Scale", minimum=1.0, maximum=10.0, step=0.1, value=1.0)
                    num_inference_steps = gr.Slider(label="Inference Steps", minimum=1, maximum=40, step=1, value=4)
                    height = gr.Slider(label="Height", minimum=256, maximum=2048, step=8, value=1024)
                    width = gr.Slider(label="Width", minimum=256, maximum=2048, step=8, value=1024)

            with gr.Column():
                result = gr.Image(label="Output Image", interactive=False)
                prompt_preview = gr.Textbox(label="Processed Prompt", interactive=False)
                create_video_button = gr.Button("🎥 Create Video Between Images", variant="secondary", visible=False)
                with gr.Group(visible=False) as video_group:
                    video_output = gr.Video(label="Generated Video", show_download_button=True, autoplay=True)
                    
    inputs = [
        image, light_type, light_direction, light_intensity, prompt,
        seed, randomize_seed, true_guidance_scale, num_inference_steps, height, width, prev_output
    ]
    outputs = [result, seed, prompt_preview]

    # Reset behavior
    reset_btn.click(
        fn=reset_all,
        inputs=None,
        outputs=[light_type, light_direction, light_intensity, prompt, is_reset],
        queue=False
    ).then(fn=end_reset, inputs=None, outputs=[is_reset], queue=False)

    # Manual generation with video button visibility control
    def infer_and_show_video_button(*args):
        result_img, result_seed, result_prompt = infer_relight(*args)
        # Show video button if we have both input and output images
        show_button = args[0] is not None and result_img is not None
        return result_img, result_seed, result_prompt, gr.update(visible=show_button)
    
    run_event = run_btn.click(
        fn=infer_and_show_video_button, 
        inputs=inputs, 
        outputs=outputs + [create_video_button]
    )

    # Video creation
    create_video_button.click(
        fn=lambda: gr.update(visible=True), 
        outputs=[video_group],
        api_name=False
    ).then(
        fn=create_video_between_images,
        inputs=[image, result, prompt_preview],
        outputs=[video_output],
        api_name=False
    )

    # Examples - You'll need to add your own example images
    gr.Examples(
        examples=[
            [None, "soft_window", "side", "soft", "", 0, True, 1.0, 4, 1024, 1024],
            [None, "golden_hour", "front", "medium", "", 0, True, 1.0, 4, 1024, 1024],
            [None, "dramatic", "side", "strong", "", 0, True, 1.0, 4, 1024, 1024],
            [None, "neon", "front", "medium", "", 0, True, 1.0, 4, 1024, 1024],
            [None, "candlelight", "front", "soft", "", 0, True, 1.0, 4, 1024, 1024],
        ],
        inputs=[image, light_type, light_direction, light_intensity, prompt,
                seed, randomize_seed, true_guidance_scale, num_inference_steps, height, width],
        outputs=outputs,
        fn=infer_relight,
        cache_examples="lazy",
        elem_id="examples"
    )
    
    # Image upload triggers dimension update and control reset
    image.upload(
        fn=update_dimensions_on_upload,
        inputs=[image],
        outputs=[width, height]
    ).then(
        fn=reset_all,
        inputs=None,
        outputs=[light_type, light_direction, light_intensity, prompt, is_reset],
        queue=False
    ).then(
        fn=end_reset, 
        inputs=None, 
        outputs=[is_reset], 
        queue=False
    )


    # Live updates
    def maybe_infer(is_reset, progress=gr.Progress(track_tqdm=True), *args):
        if is_reset:
            return gr.update(), gr.update(), gr.update(), gr.update()
        else:
            result_img, result_seed, result_prompt = infer_relight(*args)
            # Show video button if we have both input and output
            show_button = args[0] is not None and result_img is not None
            return result_img, result_seed, result_prompt, gr.update(visible=show_button)

    control_inputs = [
        image, light_type, light_direction, light_intensity, prompt,
        seed, randomize_seed, true_guidance_scale, num_inference_steps, height, width, prev_output
    ]
    control_inputs_with_flag = [is_reset] + control_inputs

    for control in [light_type, light_direction, light_intensity]:
        control.input(fn=maybe_infer, inputs=control_inputs_with_flag, outputs=outputs + [create_video_button])
    
    # prompt.change(fn=maybe_infer, inputs=control_inputs_with_flag, outputs=outputs + [create_video_button])
    
    run_event.then(lambda img, *_: img, inputs=[result], outputs=[prev_output])

demo.launch()