Cannot reproduce results from paper with pre-trained model

researchmm / AOT-GAN-for-Inpainting

[TVCG'2023] AOT-GAN for High-Resolution Image Inpainting (codebase for image inpainting)

https://arxiv.org/abs/2104.01431

Apache License 2.0

425 stars 68 forks source link

Cannot reproduce results from paper with pre-trained model #3

Open clementabary opened 3 years ago

clementabary commented 3 years ago

Hi !

Thanks for your work, the demo code is neat 🤗 I tried the celebahq pre-trained models on images and noticed it doesn't really "complete" parts of the face when you erase them completely (such as eyes, eyebrows, mouth). I tried the same masks you showed in your paper and could not reach the same quality.

This is best I could get for example : am I missing something ? test_masked test_pred

roimulia2 commented 3 years ago

Seeing the same results as well, is the celebahq model is broken?