Flux Kontext Makeup Remover
Makeup-removal LoRA for FLUX.1 Kontext trained on synthetic data: AR makeup effects and filters turned AI-generated portraits into 70 pairs that differ only in makeup, and 59 passed review. Published on Civitai, where it has passed 2,000 downloads.
Why synthetic data
A makeup-removal model needs image pairs of the same person, under the same lighting and framing, that differ only in makeup. Such pairs are hard to shoot and hard to find in public datasets, so I synthesized the paired data from scratch and reused the beauty AR effects I had built in earlier projects as data-generation tools.
Synthetic data workflow
- Portraits: Generated 70 portraits with AI. About 80% are Asian, and lighting, background, hair and accessories differ from image to image.
- Pairs from AR effects: Combined makeup effects I had developed in Effect House, Lens Studio and Spark AR for earlier beauty projects with filters from apps such as SNOW and Meitu. The effects either removed or applied makeup, so each pair keeps the same face and framing and differs only in makeup.
- Input and target: The made-up image is the input (768×768) and the bare-faced image is the target (1024×1024).
- Review: Reviewed the 70 pairs and kept 59 for training.
- Training: Trained a FLUX.1 Kontext LoRA with AI-Toolkit.
Results
A Kontext LoRA learns the edit, removing the makeup, rather than specific faces, so it works on other ethnicities even though the data is mostly Asian.
Published on Civitai, with 2,016 downloads and 203 likes as of November 9, 2026. Put Remove makeup of this person in the prompt and use a LoRA strength of 1.
The training data grid below shows 6 of the 59 pairs, two images per pair with the input on the left and the target on the right. The full dataset is not published.
gallery (21)
Training data · input → target (6 of 59 pairs)












Results · Civitai examples








