Anime Censorship Removal LoRa for Qwen Image Edit

Hello everyone!

Model Keyword: remove censor

The Problem

Recently, I encountered an issue where image editing models completely fail to understand censorship. They literally struggle to remove even simple bar censorship.

The Solution

To address this, I created a LoRa for anime images designed for Qwen Image Edit 2511. It is already showing progress and can successfully handle light to medium mosaic censorship, as well as bar censorship.

This is an early beta version. However, it is clear that expanding the dataset and increasing the training time will yield solid results. Despite being in beta, this version is already usable for light censorship.

Known Limitations and Workaround

The main issue lies with heavy mosaic censorship: the LoRa struggles to properly recreate normal-looking genitals. The solution to add the missing anatomical details is to use inpainting after applying the LoRa. This approach works significantly better than applying inpainting directly over the original mosaic censorship.

## Testing Environment

All of my tests were conducted using NSFW v23 AIO: https://huggingface.co/Phr00t/Qwen-Image-Edit-Rapid-AIO.

The results on a standard diffusion model are currently worse, which is to be expected since the LoRa is still in early beta.

Training Details

* Dataset: 97 images

* Training Steps: 5000

* LoRa Rank: 32

* Learning Rate: 0.0001

Conclusion

In summary, this current version, when paired with Phr00t's NSFW v23 AIO, is ready to be used for restoring soft and medium mosaic censoring or bar censorship, especially if there are visable some details.