🪄 EditOnline

A WebUI for image editing, support mask-based and instruction-based editing.

🐈 Features

1. MORE CONSISTENT to use our pretrained LoRA.
2. Both local deployed model and API is supported.
3. Both instruction and mask editing is supported.
4. Design your own masks and save it.

Quick Start

Python 3.14, diffusers 0.37.1 and torch 2.10 are preferred.

• Environment Configurations: This project use uv to build the environment. The configuration file have been provided, see pyproject.toml for more details. Run

uv sync
# Or uv sync --index-url other-source to enable synchronizing the environment from the specific source.

and the environment will be created in .venv by default. To modify the path of the environment, specify the env variable UV_PROJECT_ENVIRONMENT to your prefer path.

• DiT and Pre-trained weights: The base model is QwenImage-Edit-2511, we provide a pre-trained LoRA for mask-based image editing. The weights will be released.

• To modify the path which the model will be loaded from, you can look config for more details.

• External repo: The mask generator depends to external repository, SAM is used by default to generate more accurate mask. You add another segmentation models in mask_generator/model_hub.

• 🪄 START: Just run

uv run uvicorn editonline.app:app --host 127.0.0.1 --port 8000

and the server will run locally on port 8000, open the website and you will get the following panel:

Then you can upload your own image and draw/auto-segment/semi-segment a mask for localization, then tell the pipeline how to edit via an edit prompt on the right.

• Draw a mask: the panel support to draw a mask via mouse brush
• Auto-segment: invoke a segmentation model (default SAM) to generate a set of segmentations
• Semi-segment: modify a existing mask generated by model

🤓☝️ TODO

• Release EditOnline framework
• Release mask-based editing LoRA weights
• Support API query

🤗 Contributors

• The mask generator part is finished by @Shunzi Yang