sean96t/Flux-Kontext-TryOn

Virtual Try-on using Flux Kontext LoRA

A simple and efficient virtual try-on model that can run on a single RTX 4090 with 1184x864 resolution.

Online Demo:

Installation

Recommend creating a virtual env with python 3.11, e.g. with conda.

Tested on CUDA 12.8 and so will attempt to install corresponding pytorch version

pip install -r requirements.txt

Run

Clone Github repo @ FluxKontext_Tryon

python app.py

Note

We quantize the DiT to 8-bit and use torch.compile by default so first run will be slightly longer.

We also quantize the T5 text encoder to 8-bit precision as well.

Users who have access to more compute can modify the code to enable higher precision.

Safety

The model may generate some NSFW images depending on the clothes provided. Thus, we have implemented a safety checker using Qwen3-VL-8B.

NSFW images detected will be replaced by a black image.

Acknowledgement

Our code is modified based on Diffusers. We adopt Flux.1-Kontext-dev as the base model. We also use Qwen/Qwen3-VL-8B-Instruct in our code base.