WeVi's picture
Create app.py
af56fc4 verified
import gradio as gr
import numpy as np
import random
import torch
import spaces
import os
import json
from PIL import Image
from diffusers import QwenImageEditPipeline, FlowMatchEulerDiscreteScheduler
from huggingface_hub import InferenceClient
import math
# Assuming optimization.py and qwenimage/ are in the same directory
from optimization import optimize_pipeline_
from qwenimage.pipeline_qwen_image_edit import QwenImageEditPipeline as QwenImageEditPipelineCustom
from qwenimage.transformer_qwen_image import QwenImageTransformer2DModel
from qwenimage.qwen_fa3_processor import QwenDoubleStreamAttnProcessorFA3
# --- prompt enhancement using hugging face inferenceclient ---
def polish_prompt_hf(original_prompt, system_prompt):
"""
Rewrites the prompt using a Hugging Face InferenceClient.
"""
api_key = os.environ.get("HF_TOKEN") # Changed to HF_TOKEN as per common practice
if not api_key:
print("Warning: HF_TOKEN not set. Falling back to original prompt.")
return original_prompt
try:
client = InferenceClient(
provider="cerebras",
api_key=api_key,
)
messages = [
{"role": "system", "content": system_prompt},
{"role": "user", "content": original_prompt}
]
completion = client.chat.completions.create(
model="qwen/qwen3-235b-a22b-instruct-2507",
messages=messages,
)
result = completion.choices[0].message.content
if '{"rewritten"' in result:
try:
result = result.replace('```json', '').replace('```', '')
result_json = json.loads(result)
polished_prompt = result_json.get('rewritten', result)
except Exception: # Catch broader exception for JSON parsing
polished_prompt = result
else:
polished_prompt = result
polished_prompt = polished_prompt.strip().replace("\n", " ")
return polished_prompt
except Exception as e: # Catch broader exception for API calls
print(f"Error during API call to Hugging Face: {e}")
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.
'''
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"
# Scheduler configuration for lightning
scheduler_config = {
"base_image_seq_len": 256,
"base_shift": math.log(3),
"invert_sigmas": False, # Corrected boolean case
"max_image_seq_len": 8192,
"max_shift": math.log(3),
"num_train_timesteps": 1000,
"shift": 1.0,
"shift_terminal": None, # Corrected None case
"stochastic_sampling": False, # Corrected boolean case
"time_shift_type": "exponential",
"use_beta_sigmas": False, # Corrected boolean case
"use_dynamic_shifting": True, # Corrected boolean case
"use_exponential_sigmas": False, # Corrected boolean case
"use_karras_sigmas": False, # Corrected boolean case
}
# Initialize scheduler with lightning config
scheduler = FlowMatchEulerDiscreteScheduler.from_config(scheduler_config)
# Load the edit pipeline with lightning scheduler
pipe = QwenImageEditPipelineCustom.from_pretrained( # Corrected class name
"qwen/qwen-image-edit",
scheduler=scheduler,
torch_dtype=dtype
).to(device)
# Load lightning LoRA weights for acceleration
try:
pipe.load_lora_weights(
"lightx2v/qwen-image-lightning",
weight_name="qwen-image-lightning-8steps-v1.1.safetensors"
)
pipe.fuse_lora()
print("Successfully loaded lightning LoRA weights")
except Exception as e: # Catch broader exception
print(f"Warning: Could not load lightning LoRA weights: {e}")
print("Continuing with base model...")
# Apply the same optimizations from the first version
pipe.transformer.__class__ = QwenImageTransformer2DModel # Corrected class name
pipe.transformer.set_attn_processor(QwenDoubleStreamAttnProcessorFA3()) # Corrected class name
# --- Ahead-of-time compilation ---
# It's important that the dummy image for optimization has the expected dimensions (e.g., 1024x1024)
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 ---
spaces.gpu(duration=60)
def infer(
image,
prompt,
seed=42,
randomize_seed=False, # Corrected boolean case
true_guidance_scale=1.0,
num_inference_steps=8, # Default to 8 steps for fast inference
rewrite_prompt=True, # Corrected boolean case
output_size="Original (1024x1024)", # New parameter for output size
progress=gr.Progress(track_tqdm=True), # Corrected class name
):
"""
Generates an edited image using the Qwen-Image-Edit pipeline with lightning acceleration,
and optionally resizes the output.
"""
negative_prompt = " "
if randomize_seed:
seed = random.randint(0, max_seed)
generator = torch.Generator(device=device).manual_seed(seed) # Corrected class name
print(f"Original prompt: '{prompt}'")
print(f"Negative prompt: '{negative_prompt}'")
print(f"Seed: {seed}, Steps: {num_inference_steps}, Guidance: {true_guidance_scale}")
if rewrite_prompt:
prompt = polish_prompt(prompt, image)
print(f"Rewritten prompt: {prompt}")
try:
images = pipe(
image,
prompt=prompt,
negative_prompt=negative_prompt,
num_inference_steps=num_inference_steps,
generator=generator,
true_cfg_scale=true_guidance_scale,
num_images_per_prompt=1
).images
output_image = images[0]
# Post-processing: Resize if a different output size is selected
if output_size != "Original (1024x1024)":
try:
if output_size == "Small (512x512)":
target_size = (512, 512)
elif output_size == "Medium (768x768)":
target_size = (768, 768)
elif output_size == "Large (1536x1536)":
target_size = (1536, 1536)
else: # Custom size, parse it from "Custom (WxH)"
width, height = map(int, output_size.split('(')[1][:-1].split('x'))
target_size = (width, height)
output_image = output_image.resize(target_size, Image.LANCZOS) # Use LANCZOS for high quality down/upscaling
print(f"Resized output image to: {target_size[0]}x{target_size[1]}")
except Exception as resize_e:
print(f"Warning: Could not resize image to {output_size}: {resize_e}")
print("Returning original size image.")
return output_image, seed
except Exception as e:
print(f"Error during inference: {e}")
raise e
# --- Examples and UI layout ---
examples = [
# Example for demonstration, replace with actual image paths
# Ensure these paths are valid if running locally, or adjust for Hugging Face Spaces
[Image.new("RGB", (1024, 1024), color = 'red'), "Change the color to blue"],
[Image.new("RGB", (1024, 1024), color = 'green'), "Add a fluffy white cat sitting in the center"],
]
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">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 [Qwen-Image-Lightning](https://huggingface.co/lightx2v/qwen-image-lightning) LoRA for accelerated inference.
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=True,
type="pil"
)
result = gr.Image(
label="Result",
show_label=True,
type="pil"
)
with gr.Row():
prompt = gr.Text(
label="Edit Instruction",
show_label=False,
placeholder="Describe the edit instruction (e.g., 'replace the background with a sunset', 'add a red hat', 'remove the person')",
container=False,
)
run_button = gr.Button("Edit!", 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)
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=4,
maximum=28,
step=1,
value=8
)
rewrite_prompt = gr.Checkbox(
label="Enhance Prompt (using HF Inference)",
value=True
)
# New dropdown for output image size
output_size = gr.Dropdown(
label="Output Image Size",
choices=["Original (1024x1024)", "Small (512x512)", "Medium (768x768)", "Large (1536x1536)"],
value="Original (1024x1024)"
)
gr.Examples(examples=examples, inputs=[input_image, prompt], outputs=[result, seed], fn=infer, cache_examples=False) # Changed to use the new example inputs/outputs
gr.on(
triggers=[run_button.click, prompt.submit],
fn=infer,
inputs=[
input_image,
prompt,
seed,
randomize_seed,
true_guidance_scale,
num_inference_steps,
rewrite_prompt,
output_size, # Added output_size to inputs
],
outputs=[result, seed],
)
if __name__ == "__main__":
demo.launch()