Model generalization ability？

[yuchen-ji]

Can the model in this article be generalized on images outside the dataset, such as completing a masked chair image?

[idealwhite]

It can not. The provided weights are trained on a limited small dataset. But the method could be used for open-domain data. For example, you can replace the image encoder with CLIP and train it on a large image-text dataset. Besides, existing diffusion models like GLIDE and StableDiffusion provide better paint ability. If your machine can not run a diffusion model, you may use this TDANet network.

[yuchen-ji]

I see. Thank you very much for your work and suggestions!