erickillian / tattoo-remover

Tattoo Remover that uses UNet and MSE Loss
MIT License
3 stars 1 forks source link

Room for improvement #1

Open kexul opened 8 months ago

kexul commented 8 months ago

Huge thanks for open souce! 👍 👍 👍 The result is amazing! But some of the tattoo is not removed: 图片 图片

How could I improve the result?

erickillian commented 8 months ago

Hi Kexul! Thanks for checking out my project.

I am still working on trying to improve the model as well as the dataset which could improve the results. If you are interested in helping I think the biggest thing would be to help with the dataset collection of paired tattoo / no tattoo photos.

kexul commented 8 months ago

Is there some way to use the result or model from stable diffusion? I've got decent result by inpainting the tatto area using stable diffusion model. Original image: 图片 By tattoo remover: 无标题 By stable diffusion inpainting: 图片

kexul commented 8 months ago

Here is the tool I used for inpainting: https://github.com/Sanster/IOPaint. The disadvantage of the inpainting method is that the mask need to be drawn manually, which is very tedious when the tattoo area is very large, for example the naked Justin Biber.

kexul commented 8 months ago

Could we use the reuslt of the tattoo-remover to be the mask of inpainting? 🤔

kexul commented 8 months ago

Yes, the mask can easily be obtained by 图片 , I may need to binary the mask, do some morphology operations then.

kexul commented 8 months ago

Oh.. the result is not as good as tattoo remover, when use a mask.

图片