question！

[bxhsort]

Hi, I want to ask you a question, is the process the same as the diffusion idea for adding convolutional layers to retain more information? Algorithm 1 is for denoising, and algorithm 2 is for denoising the reconstructed image! I have a question about this place and I hope you can answer it. Thank you very much

[iFighting]

Hi, I want to ask you a question, is the process the same as the diffusion idea for adding convolutional layers to retain more information? Algorithm 1 is for denoising, and algorithm 2 is for denoising the reconstructed image! I have a question about this place and I hope you can answer it. Thank you very much

@bxhsort Actually, it is not the case, The algorithm 1 and 2 is for the var image tokenizer encoding & decoding setting, not the image generation process. The image generation is done by the autoregressive process.

[bxhsort]

Hi, I want to ask you a question, is the process the same as the diffusion idea for adding convolutional layers to retain more information? Algorithm 1 is for denoising, and algorithm 2 is for denoising the reconstructed image! I have a question about this place and I hope you can answer it. Thank you very much

@bxhsort Actually, it is not the case, The algorithm 1 and 2 is for the var image tokenizer encoding & decoding setting, not the image generation process. The image generation is done by the autoregressive process.

Thank you for your answer. My understanding is that algorithms 1 and 2 provide GT for the subsequent autoregressive training. Is this process to better train the token and its Embedding, which is similar to the training of tokens in nlp? The process of stage2 is an autoregressive process, which is also a process of inference. In algorithm 1, after you extract features with convolution layer, you subtract them with image feature f. I have some doubts about this place, and I would like to ask why you can save information well by doing this. Looking forward to your reply, thank you very much!