import gradio as gr import numpy as np import random import torch import spaces from PIL import Image import torch import math from optimization import optimize_pipeline_ from qwenimage.pipeline_qwen_image_edit import QwenImageEditPipeline from qwenimage.transformer_qwenimage import QwenImageTransformer2DModel from huggingface_hub import InferenceClient import math # --- 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) # --- Model Loading --- dtype = torch.bfloat16 device = "cuda" if torch.cuda.is_available() else "cpu" pipe = QwenImageEditPipeline.from_pretrained("Qwen/Qwen-Image-Edit", torch_dtype=dtype).to(device) pipe.transformer.__class__ = QwenImageTransformer2DModel # --- Ahead-of-time compilation --- optimize_pipeline_(pipe, image=Image.new("RGB", (1024, 1024)), prompt="prompt") # --- UI Constants and Helpers --- MAX_SEED = np.iinfo(np.int32).max # --- Main Inference Function (with hardcoded negative prompt) --- @spaces.GPU(duration=120) def infer( image, prompt, seed=42, randomize_seed=False, true_guidance_scale=4.0, num_inference_steps=50, rewrite_prompt=True, 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) print(f"Calling pipeline with prompt: '{prompt}'") print(f"Negative Prompt: '{negative_prompt}'") print(f"Seed: {seed}, Steps: {num_inference_steps}") if rewrite_prompt: prompt = polish_prompt(prompt, image) print(f"Rewritten Prompt: {prompt}") # Generate the image image = pipe( image, prompt=prompt, negative_prompt=negative_prompt, num_inference_steps=num_inference_steps, generator=generator, true_cfg_scale=true_guidance_scale, ).images[0] return image, seed # --- Examples and UI Layout --- examples = [] css = """ #col-container { margin: 0 auto; max-width: 1024px; } #edit_text{margin-top: -62px !important} """ with gr.Blocks(css=css) as demo: with gr.Column(elem_id="col-container"): gr.HTML('Qwen-Image Logo') gr.HTML('

Edit

', elem_id="edit_text") gr.Markdown("[Learn more](https://github.com/QwenLM/Qwen-Image) about the Qwen-Image series. Try on [Qwen Chat](https://chat.qwen.ai/), or [download model](https://huggingface.co/Qwen/Qwen-Image-Edit) to run locally with ComfyUI or diffusers.") with gr.Row(): with gr.Column(): input_image = gr.Image(label="Input Image", show_label=False, type="pil") prompt = gr.Text( label="Prompt", show_label=False, placeholder="describe the edit instruction", container=False, ) run_button = gr.Button("Edit!", variant="primary") result = gr.Image(label="Result", show_label=False, type="pil") 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=50, step=1, value=50, ) rewrite_prompt = gr.Checkbox( label="Enhance prompt (using HF Inference)", value=True ) # 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_image, prompt, # negative_prompt is no longer an input from the UI seed, randomize_seed, true_guidance_scale, num_inference_steps, rewrite_prompt, ], outputs=[result, seed], ) if __name__ == "__main__": demo.launch()