app_temp.py

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

from io import BytesIO
from PIL import Image
from diffusers import FlowMatchEulerDiscreteScheduler
from optimization_simple import optimize_pipeline_
from qwenimage.pipeline_qwenimage_edit_plus import QwenImageEditPlusPipeline
from qwenimage.transformer_qwenimage import QwenImageTransformer2DModel
from qwenimage.qwen_fa3_processor import QwenDoubleStreamAttnProcessorFA3

from huggingface_hub import InferenceClient
import math

import os
import base64
import json


SYSTEM_PROMPT = '''
# Edit Instruction Rewriter
You are a professional edit instruction rewriter. Your task is to generate a precise, concise, and visually achievable professional-level edit instruction based on the user-provided instruction and the image to be edited.  
Please strictly follow the rewriting rules below:
## 1. General Principles
- Keep the rewritten prompt **concise and comprehensive**. Avoid overly long sentences and unnecessary descriptive language.  
- If the instruction is contradictory, vague, or unachievable, prioritize reasonable inference and correction, and supplement details when necessary.  
- Keep the main part of the original instruction unchanged, only enhancing its clarity, rationality, and visual feasibility.  
- All added objects or modifications must align with the logic and style of the scene in the input images.  
- If multiple sub-images are to be generated, describe the content of each sub-image individually.  
## 2. Task-Type Handling Rules
### 1. Add, Delete, Replace Tasks
- If the instruction is clear (already includes task type, target entity, position, quantity, attributes), preserve the original intent and only refine the grammar.  
- If the description is vague, supplement with minimal but sufficient details (category, color, size, orientation, position, etc.). For example:  
    > Original: "Add an animal"  
    > Rewritten: "Add a light-gray cat in the bottom-right corner, sitting and facing the camera"  
- Remove meaningless instructions: e.g., "Add 0 objects" should be ignored or flagged as invalid.  
- For replacement tasks, specify "Replace Y with X" and briefly describe the key visual features of X.  
### 2. Text Editing Tasks
- All text content must be enclosed in English double quotes `" "`. Keep the original language of the text, and keep the capitalization.  
- Both adding new text and replacing existing text are text replacement tasks, For example:  
    - Replace "xx" to "yy"  
    - Replace the mask / bounding box to "yy"  
    - Replace the visual object to "yy"  
- Specify text position, color, and layout only if user has required.  
- If font is specified, keep the original language of the font.  
### 3. Human Editing Tasks
- Make the smallest changes to the given user's prompt.  
- If changes to background, action, expression, camera shot, or ambient lighting are required, please list each modification individually.
- **Edits to makeup or facial features / expression must be subtle, not exaggerated, and must preserve the subject's identity consistency.**
    > Original: "Add eyebrows to the face"  
    > Rewritten: "Slightly thicken the person's eyebrows with little change, look natural."
### 4. Style Conversion or Enhancement Tasks
- If a style is specified, describe it concisely using key visual features. For example:  
    > Original: "Disco style"  
    > Rewritten: "1970s disco style: flashing lights, disco ball, mirrored walls, vibrant colors"  
- For style reference, analyze the original image and extract key characteristics (color, composition, texture, lighting, artistic style, etc.), integrating them into the instruction.  
- **Colorization tasks (including old photo restoration) must use the fixed template:**  
  "Restore and colorize the old photo."  
- Clearly specify the object to be modified. For example:  
    > Original: Modify the subject in Picture 1 to match the style of Picture 2.  
    > Rewritten: Change the girl in Picture 1 to the ink-wash style of Picture 2 — rendered in black-and-white watercolor with soft color transitions.
### 5. Material Replacement
- Clearly specify the object and the material. For example: "Change the material of the apple to papercut style."
- For text material replacement, use the fixed template:
    "Change the material of text "xxxx" to laser style"
### 6. Logo/Pattern Editing
- Material replacement should preserve the original shape and structure as much as possible. For example:
   > Original: "Convert to sapphire material"  
   > Rewritten: "Convert the main subject in the image to sapphire material, preserving similar shape and structure"
- When migrating logos/patterns to new scenes, ensure shape and structure consistency. For example:
   > Original: "Migrate the logo in the image to a new scene"  
   > Rewritten: "Migrate the logo in the image to a new scene, preserving similar shape and structure"
### 7. Multi-Image Tasks
- Rewritten prompts must clearly point out which image's element is being modified. For example:  
    > Original: "Replace the subject of picture 1 with the subject of picture 2"  
    > Rewritten: "Replace the girl of picture 1 with the boy of picture 2, keeping picture 2's background unchanged"  
- For stylization tasks, describe the reference image's style in the rewritten prompt, while preserving the visual content of the source image.  
## 3. Rationale and Logic Check
- Resolve contradictory instructions: e.g., "Remove all trees but keep all trees" requires logical correction.
- Supplement missing critical information: e.g., if position is unspecified, choose a reasonable area based on composition (near subject, blank space, center/edge, etc.).
# Output Format Example
```json
{
   "Rewritten": "..."
}
'''

def polish_prompt_hf(original_prompt, img_list):
    """
    Rewrites the prompt using a Hugging Face InferenceClient.
    Supports multiple images via img_list.
    """
    # Ensure HF_TOKEN is set
    api_key = os.environ.get("HF_TOKEN")
    if not api_key:
        print("Warning: HF_TOKEN not set. Falling back to original prompt.")
        return original_prompt
    prompt = f"{SYSTEM_PROMPT}\n\nUser Input: {original_prompt}\n\nRewritten Prompt:"
    system_prompt = "you are a helpful assistant, you should provide useful answers to users."
    try:
        # Initialize the client
        client = InferenceClient(
            provider="nebius",
            api_key=api_key,
        )

        # Convert list of images to base64 data URLs
        image_urls = []
        if img_list is not None:
            # Ensure img_list is actually a list
            if not isinstance(img_list, list):
                img_list = [img_list]
            
            for img in img_list:
                image_url = None
                # If img is a PIL Image
                if hasattr(img, 'save'):  # Check if it's a PIL Image
                    buffered = BytesIO()
                    img.save(buffered, format="PNG")
                    img_base64 = base64.b64encode(buffered.getvalue()).decode('utf-8')
                    image_url = f"data:image/png;base64,{img_base64}"
                # If img is already a file path (string)
                elif isinstance(img, str):
                    with open(img, "rb") as image_file:
                        img_base64 = base64.b64encode(image_file.read()).decode('utf-8')
                    image_url = f"data:image/png;base64,{img_base64}"
                else:
                    print(f"Warning: Unexpected image type: {type(img)}, skipping...")
                    continue
                
                if image_url:
                    image_urls.append(image_url)

        # Build the content array with text first, then all images
        content = [
            {
                "type": "text",
                "text": prompt
            }
        ]
        
        # Add all images to the content
        for image_url in image_urls:
            content.append({
                "type": "image_url",
                "image_url": {
                    "url": image_url
                }
            })

        # Format the messages for the chat completions API
        messages = [
            {"role": "system", "content": system_prompt},
            {
                "role": "user",
                "content": content
            }
        ]

        # Call the API
        completion = client.chat.completions.create(
            model="Qwen/Qwen2.5-VL-72B-Instruct",
            messages=messages,
        )
        
        # Parse the response
        result = completion.choices[0].message.content
        
        # Try to extract JSON if present
        if '"Rewritten"' in result:
            try:
                # Clean up the response
                result = result.replace('```json', '').replace('```', '')
                result_json = json.loads(result)
                polished_prompt = result_json.get('Rewritten', result)
            except:
                polished_prompt = result
        else:
            polished_prompt = result
            
        polished_prompt = polished_prompt.strip().replace("\n", " ")
        return polished_prompt
        
    except Exception as e:
        print(f"Error during API call to Hugging Face: {e}")
        # Fallback to original prompt if enhancement fails
        return original_prompt 

def encode_image(pil_image):
    import io
    buffered = io.BytesIO()
    pil_image.save(buffered, format="PNG")
    return base64.b64encode(buffered.getvalue()).decode("utf-8")

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

# Scheduler configuration for Lightning
scheduler_config = {
    "base_image_seq_len": 256,
    "base_shift": math.log(3),
    "invert_sigmas": False,
    "max_image_seq_len": 8192,
    "max_shift": math.log(3),
    "num_train_timesteps": 1000,
    "shift": 1.0,
    "shift_terminal": None,
    "stochastic_sampling": False,
    "time_shift_type": "exponential",
    "use_beta_sigmas": False,
    "use_dynamic_shifting": True,
    "use_exponential_sigmas": False,
    "use_karras_sigmas": False,
}

# Initialize scheduler with Lightning config
scheduler = FlowMatchEulerDiscreteScheduler.from_config(scheduler_config)

# Load the model pipeline
pipe = QwenImageEditPlusPipeline.from_pretrained("Qwen/Qwen-Image-Edit-2509", 
                                                 scheduler=scheduler,
                                                 torch_dtype=dtype).to(device)
pipe.load_lora_weights(
        "lightx2v/Qwen-Image-Lightning", 
        weight_name="Qwen-Image-Edit-2509/Qwen-Image-Edit-2509-Lightning-4steps-V1.0-fp32.safetensors"
    )
pipe.fuse_lora()

# Apply the same optimizations from the first version
pipe.transformer.__class__ = QwenImageTransformer2DModel
pipe.transformer.set_attn_processor(QwenDoubleStreamAttnProcessorFA3())

# --- Ahead-of-time compilation ---
optimize_pipeline_(pipe, image=[Image.new("RGB", (1024, 1024)), Image.new("RGB", (1024, 1024))], prompt="prompt")

# --- UI Constants and Helpers ---
MAX_SEED = np.iinfo(np.int32).max

def use_output_as_input(output_images):
    """Convert output images to input format for the gallery"""
    if output_images is None or len(output_images) == 0:
        return []
    return output_images
    
# --- Main Inference Function (with hardcoded negative prompt) ---
@spaces.GPU(duration=40)
def infer(
    images,
    prompt,
    seed=42,
    randomize_seed=False,
    true_guidance_scale=1.0,
    num_inference_steps=4,
    height=None,
    width=None,
    rewrite_prompt=True,
    num_images_per_prompt=2,
    progress=gr.Progress(track_tqdm=True),
):
    """
    Generates an image using the local Qwen-Image diffusers pipeline.
    """
    # Hardcode the negative prompt as requested
    negative_prompt = " "
    
    if randomize_seed:
        seed = random.randint(0, MAX_SEED)

    # Set up the generator for reproducibility
    generator = torch.Generator(device=device).manual_seed(seed)
    
    # Load input images into PIL Images
    pil_images = []
    if images is not None:
        for item in images:
            try:
                if isinstance(item[0], Image.Image):
                    pil_images.append(item[0].convert("RGB"))
                elif isinstance(item[0], str):
                    pil_images.append(Image.open(item[0]).convert("RGB"))
                elif hasattr(item, "name"):
                    pil_images.append(Image.open(item.name).convert("RGB"))
            except Exception:
                continue

    if height==256 and width==256:
        height, width = None, None
    print(f"Calling pipeline with prompt: '{prompt}'")
    print(f"Negative Prompt: '{negative_prompt}'")
    print(f"Seed: {seed}, Steps: {num_inference_steps}, Guidance: {true_guidance_scale}, Size: {width}x{height}")
    if rewrite_prompt and len(pil_images) > 0:
        prompt = polish_prompt_hf(prompt, pil_images)
        print(f"Rewritten Prompt: {prompt}")
    

    # Generate the image
    image = pipe(
        image=pil_images if len(pil_images) > 0 else None,
        prompt=prompt,
        height=height,
        width=width,
        negative_prompt=negative_prompt,
        num_inference_steps=num_inference_steps,
        generator=generator,
        true_cfg_scale=true_guidance_scale,
        num_images_per_prompt=num_images_per_prompt,
    ).images

    return image, seed

# --- Examples and UI Layout ---
examples = []

css = """
#col-container {
    margin: 0 auto;
    max-width: 1024px;
}
#logo-title {
    text-align: center;
}
#logo-title img {
    width: 400px;
}
#edit_text{margin-top: -62px !important}
"""

with gr.Blocks(css=css) as demo:
    with gr.Column(elem_id="col-container"):
        gr.HTML("""
        <div id="logo-title">
            <img src="https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-Image/qwen_image_edit_logo.png" alt="Qwen-Image Edit Logo" width="400" style="display: block; margin: 0 auto;">
            <h2 style="font-style: italic;color: #5b47d1;margin-top: -27px !important;margin-left: 96px">[Plus] Fast, 8-steps with Lightning LoRA</h2>
        </div>
        """)
        gr.Markdown("""
        [Learn more](https://github.com/QwenLM/Qwen-Image) about the Qwen-Image series. 
        This demo uses the new [Qwen-Image-Edit-2509](https://huggingface.co/Qwen/Qwen-Image-Edit-2509) with the [Qwen-Image-Lightning v2](https://huggingface.co/lightx2v/Qwen-Image-Lightning) LoRA + [AoT compilation & FA3](https://huggingface.co/blog/zerogpu-aoti) for accelerated inference.
        Try on [Qwen Chat](https://chat.qwen.ai/), or [download model](https://huggingface.co/Qwen/Qwen-Image-Edit-2509) to run locally with ComfyUI or diffusers.
        """)
        with gr.Row():
            with gr.Column():
                input_images = gr.Gallery(label="Input Images", 
                                          show_label=False, 
                                          type="pil", 
                                          interactive=True)

            # result = gr.Image(label="Result", show_label=False, type="pil")
            result = gr.Gallery(label="Result", show_label=False, type="pil")
        with gr.Row():
            prompt = gr.Text(
                    label="Prompt",
                    show_label=False,
                    placeholder="describe the edit instruction",
                    container=False,
            )
            run_button = gr.Button("Edit!", variant="primary")

        with gr.Accordion("Advanced Settings", open=False):
            # Negative prompt UI element is removed here

            seed = gr.Slider(
                label="Seed",
                minimum=0,
                maximum=MAX_SEED,
                step=1,
                value=0,
            )

            randomize_seed = gr.Checkbox(label="Randomize seed", value=True)

            with gr.Row():

                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="Number of inference steps",
                    minimum=1,
                    maximum=40,
                    step=1,
                    value=4,
                )
                
                height = gr.Slider(
                    label="Height",
                    minimum=256,
                    maximum=2048,
                    step=8,
                    value=None,
                )
                
                width = gr.Slider(
                    label="Width",
                    minimum=256,
                    maximum=2048,
                    step=8,
                    value=None,
                )

                num_images = gr.Slider(
                    label="Num Images per Prompt",
                    minimum=1,
                    maximum=4,
                    step=1,
                    value=1,
                )
                
                
            rewrite_prompt = gr.Checkbox(label="Rewrite prompt (being fixed)", value=False)

        # gr.Examples(examples=examples, inputs=[prompt], outputs=[result, seed], fn=infer, cache_examples=False)

    gr.on(
        triggers=[run_button.click, prompt.submit],
        fn=infer,
        inputs=[
            input_images,
            prompt,
            seed,
            randomize_seed,
            true_guidance_scale,
            num_inference_steps,
            height,
            width,
            rewrite_prompt,
            num_images
        ],
        outputs=[result, seed],
    )

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