Qwen-Image-Edit: Alibaba’s Open‑Source State‑of‑the‑Art Image Editing Model

Model Overview

Qwen‑Image‑Edit is an image‑editing model built on the 20‑billion‑parameter Qwen‑Image foundation model. It adds two capabilities: semantic editing (changing image content such as style, rotation, artistic style) and appearance editing (local pixel‑level modifications such as removing objects while preserving shadows, recoloring clothing without affecting faces).

Dual‑Channel Architecture

Semantic Understanding Path : uses the Qwen2.5‑VL model to parse image semantics (e.g., “this is a dog”, “front view”).

Appearance Encoding Path : employs a VAE encoder to extract pixel‑level features (edges, colors, lighting).

The parallel paths allow simultaneous reasoning about content intent and visual form. Example: rotating a car 180° produces a realistic rear view; adding a billboard automatically generates appropriate reflections.

Image‑Text Editing

Unlike most image models that distort fonts, Qwen‑Image‑Edit can add, delete, or modify specific characters while preserving original font, size, and style for both Chinese and English.

Demonstrated Use Cases

Character image editing : edited a mascot across 16 MBTI personality types, varying emotion, pose, and art style while keeping identity consistent.

Viewpoint control : 90°/180° rotations that reconstruct back‑view geometry, useful for product visualization and AR/VR.

Style transfer : conversion to specific artistic styles (e.g., Ghibli) without distortion or boundary artifacts.

Flaw removal : precise removal of stray hairs or cluttered backgrounds while preserving shadows and edges.

Background/clothing replacement : changes shirt color or background without color bleed.

Multi‑step correction : chain editing of a calligraphy example—first generate a flawed piece, then iteratively refine overall structure and individual strokes based on region‑specific prompts.

Access and Inference

Model weights are publicly available on HuggingFace at https://huggingface.co/Qwen/Qwen-Image-Edit.

import os
from PIL import Image
import torch
from diffusers import QwenImageEditPipeline

pipeline = QwenImageEditPipeline.from_pretrained("Qwen/Qwen-Image-Edit")
print("pipeline loaded")
pipeline.to(torch.bfloat16)
pipeline.to("cuda")
pipeline.set_progress_bar_config(disable=None)

image = Image.open("./input.png").convert("RGB")
prompt = "Change the rabbit's color to purple, with a flash light background."
inputs = {
    "image": image,
    "prompt": prompt,
    "generator": torch.manual_seed(0),
    "true_cfg_scale": 4.0,
    "negative_prompt": " ",
    "num_inference_steps": 50,
}
with torch.inference_mode():
    output = pipeline(**inputs)
    output_image = output.images[0]
    output_image.save("output_image_edit.png")
    print("image saved at", os.path.abspath("output_image_edit.png"))

Benchmark

Evaluation on public image‑editing datasets shows performance that surpasses other base models, demonstrating reliable, controllable, and reversible local edits, strict adherence to user instructions, accurate region identification, correct text handling, and support for continuous refinements.