app.py

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


from diffusers.utils import load_image

from diffusers import QwenImageControlNetModel, QwenImageControlNetInpaintPipeline

import math
from huggingface_hub import InferenceClient

from PIL import Image

# Set environment variable for parallel loading
# os.environ["HF_ENABLE_PARALLEL_LOADING"] = "YES"

# --- Prompt Enhancement using Hugging Face InferenceClient ---
def polish_prompt_hf(original_prompt, system_prompt):
    """
    Rewrites the prompt using a Hugging Face InferenceClient.
    """
    # 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

    try:
        # Initialize the client
        client = InferenceClient(
            provider="cerebras",
            api_key=api_key,
        )

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

        # Call the API
        completion = client.chat.completions.create(
            model="Qwen/Qwen3-235B-A22B-Instruct-2507",
            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 polish_prompt(prompt, img):
    """
    Main function to polish prompts for image editing using HF inference.
    """
    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**. Avoid overly long sentences and reduce unnecessary descriptive language.  
- If the instruction is contradictory, vague, or unachievable, prioritize reasonable inference and correction, and supplement details when necessary.  
- Keep the core intention 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 edited input image's overall scene.  
## 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 " ". Do not translate or alter the original language of the text, and do not change the capitalization.  
- **For text replacement tasks, always use the fixed template:**
    - Replace "xx" to "yy".  
    - Replace the xx bounding box to "yy".  
- If the user does not specify text content, infer and add concise text based on the instruction and the input image's context. For example:  
    > Original: "Add a line of text" (poster)  
    > Rewritten: "Add text "LIMITED EDITION" at the top center with slight shadow"  
- Specify text position, color, and layout in a concise way.  
### 3. Human Editing Tasks
- Maintain the person's core visual consistency (ethnicity, gender, age, hairstyle, expression, outfit, etc.).  
- If modifying appearance (e.g., clothes, hairstyle), ensure the new element is consistent with the original style.  
- **For expression changes, they must be natural and subtle, never exaggerated.**  
- If deletion is not specifically emphasized, the most important subject in the original image (e.g., a person, an animal) should be preserved.
    - For background change tasks, emphasize maintaining subject consistency at first.  
- Example:  
    > Original: "Change the person's hat"  
    > Rewritten: "Replace the man's hat with a dark brown beret; keep smile, short hair, and gray jacket unchanged"  
### 4. Style Transformation or Enhancement Tasks
- If a style is specified, describe it concisely with key visual traits. For example:  
    > Original: "Disco style"  
    > Rewritten: "1970s disco: flashing lights, disco ball, mirrored walls, colorful tones"  
- If the instruction says "use reference style" or "keep current style," analyze the input image, extract main features (color, composition, texture, lighting, art style), and integrate them concisely.  
- **For coloring tasks, including restoring old photos, always use the fixed template:** "Restore old photograph, remove scratches, reduce noise, enhance details, high resolution, realistic, natural skin tones, clear facial features, no distortion, vintage photo restoration"  
- If there are other changes, place the style description at the end.
## 3. Rationality and Logic Checks
- Resolve contradictory instructions: e.g., "Remove all trees but keep all trees" should be logically corrected.  
- Add missing key information: if position is unspecified, choose a reasonable area based on composition (near subject, empty space, center/edges).  
# Output Format
Return only the rewritten instruction text directly, without JSON formatting or any other wrapper.
'''
    
    # Note: We're not actually using the image in the HF version, 
    # but keeping the interface consistent
    full_prompt = f"{SYSTEM_PROMPT}\n\nUser Input: {prompt}\n\nRewritten Prompt:"
    
    return polish_prompt_hf(full_prompt, SYSTEM_PROMPT)


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

# --- Helper functions for reuse feature ---
def clear_result():
    """Clears the result image."""
    return gr.update(value=None)

def use_output_as_input(output_image):
    """Sets the generated output as the new input image."""
    if output_image is not None:
        return gr.update(value=output_image[1])
    return gr.update()


base_model = "Qwen/Qwen-Image"
controlnet_model = "InstantX/Qwen-Image-ControlNet-Inpainting"

controlnet = QwenImageControlNetModel.from_pretrained(controlnet_model, torch_dtype=torch.bfloat16)

pipe = QwenImageControlNetInpaintPipeline.from_pretrained(
    base_model, controlnet=controlnet, torch_dtype=torch.bfloat16
)
pipe.to("cuda")


@spaces.GPU(duration=120)
def infer(edit_images, 
          prompt, 
          negative_prompt=" ", 
          seed=42, 
          randomize_seed=False, 
          strength=1.0, 
          num_inference_steps=30, 
          true_cfg_scale=4.0,
          rewrite_prompt=True,
          progress=gr.Progress(track_tqdm=True)):
    
    image = edit_images["background"]
    mask = edit_images["layers"][0]
    
    if randomize_seed:
        seed = random.randint(0, MAX_SEED)

    if rewrite_prompt:
        prompt = polish_prompt(prompt, image)
        print(f"Rewritten Prompt: {prompt}")
    
    # Generate image using Qwen pipeline
    result_image = pipe(
        prompt=prompt,
        negative_prompt=negative_prompt,
        control_image=image,
        control_mask=mask,
        controlnet_conditioning_scale=strength,
        num_inference_steps=num_inference_steps,
        true_cfg_scale=true_cfg_scale,
        generator=torch.Generator(device="cuda").manual_seed(seed)
    ).images[0]
    
    return [image, result_image], seed
    
examples = [
    "change the hat to red",
    "make the background a beautiful sunset",
    "replace the object with a flower vase",
]

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, theme=gr.themes.Citrus()) as demo:
    gr.HTML("<h1 style='text-align: center'>Qwen-Image with InstantX Inpainting ControlNet</style>")
    gr.Markdown(
        "Generate images with the [InstantX/Qwen-Image-ControlNet-Inpainting](https://huggingface.co/InstantX/Qwen-Image-ControlNet-Inpainting) that takes depth, pose and canny conditionings"
    )
    with gr.Row():
        with gr.Column():
            edit_image = gr.ImageEditor(
                label='Upload and draw mask for inpainting',
                type='pil',
                sources=["upload", "webcam"],
                image_mode='RGB',
                layers=False,
                brush=gr.Brush(colors=["#FFFFFF"], color_mode="fixed"),
                height=600
            )
            prompt = gr.Text(
                label="Prompt",
                show_label=False,
                max_lines=1,
                placeholder="Enter your prompt (e.g., 'change the hat to red')",
                container=False,
            )
            negative_prompt = gr.Text(
                label="Negative Prompt",
                show_label=True,
                max_lines=1,
                placeholder="Enter what you don't want (optional)",
                container=False,
                value="",
                visible=False
            )
            run_button = gr.Button("Run")
            
        with gr.Column():
            result = gr.ImageSlider(label="Result", show_label=False, interactive=False)
            use_as_input_button = gr.Button("🔄 Use as Input Image", visible=False, variant="secondary")
    
    with gr.Accordion("Advanced Settings", open=False):
        
        seed = gr.Slider(
            label="Seed",
            minimum=0,
            maximum=MAX_SEED,
            step=1,
            value=42,
        )
        
        randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
        
        
        with gr.Row():
            strength = gr.Slider(
                label="Conditioning Scale",
                minimum=0.0,
                maximum=1.0,
                step=0.1,
                value=1.0,
                info="Controls how much the inpainted region should change"
            )
            
            true_cfg_scale = gr.Slider(
                label="True CFG Scale",
                minimum=1.0,
                maximum=10.0,
                step=0.5,
                value=4.0,
                info="Classifier-free guidance scale"
            )

            num_inference_steps = gr.Slider(
                label="Number of inference steps",
                minimum=1,
                maximum=50,
                step=1,
                value=30,
            )
        
        rewrite_prompt = gr.Checkbox(
            label="Enhance prompt (using HF Inference)", 
            value=True
        )

    # Event handlers for reuse functionality
    use_as_input_button.click(
        fn=use_output_as_input,
        inputs=[result],
        outputs=[edit_image],
        show_api=False
    )

    # Main generation pipeline with result clearing and button visibility
    gr.on(
        triggers=[run_button.click, prompt.submit],
        fn=clear_result,
        inputs=None,
        outputs=result,
        show_api=False
    ).then(
        fn=infer,
        inputs=[edit_image, prompt, negative_prompt, seed, randomize_seed, strength, num_inference_steps, true_cfg_scale, rewrite_prompt],
        outputs=[result, seed]
    ).then(
        fn=lambda: gr.update(visible=True),
        inputs=None,
        outputs=use_as_input_button,
        show_api=False
    )

demo.launch()