Some Questions and Comments

[tom99763]

1. Do you consider that instead of the feature map from CNN, using vector-quantized AE (VQVAE) for the future work? I think the result will be surprised due to its feature compression and sampleable properties for image-to-image translation task.

2. It seems like the input-output pixel correlation largely impacts the translation result during early training process (multimodal translation or Animal-to-Human translation). Instead of predicting all at ones, two stage model (first contour, next texture) may improves the result.

Thank you

[taesungp]

Hello, thanks for suggestions.

1. I think incorporating VQVAE can be a good direction, particularly for saving compute.
2. It may, especially if we go to higher resolution. But two-stage approaches are also more cumbersome to train.