Hi, thank you for the excellent work!
I noticed that BAR uses FSQ for quantization, with binary levels applied independently across channels. This seems closely related to LFQ, as used in MAGVIT-v2. My current understanding is that one notable difference is that LFQ typically introduces an entropy regularization loss to encourage codebook utilization and prevent codebook collapse, whereas BAR does not appear to require such a loss.
Could you clarify the main differences between the FSQ formulation used in BAR and LFQ? In particular, are there any important architectural or optimization differences beyond the use of entropy regularization?
I am also curious about why entropy loss is not necessary in BAR. One possibility is that the encoder distillation or alignment objective with the frozen SigLIP2 encoder implicitly encourages diverse and informative code usage, thereby serving a role similar to entropy regularization. Another possibility is that entropy regularization becomes less important when the quantization dimensionality—and therefore the effective codebook size—is relatively modest, such as in the 16-bit setting.
I would greatly appreciate any clarification or insight from the authors. Thank you!
Hi, thank you for the excellent work!
I noticed that BAR uses FSQ for quantization, with binary levels applied independently across channels. This seems closely related to LFQ, as used in MAGVIT-v2. My current understanding is that one notable difference is that LFQ typically introduces an entropy regularization loss to encourage codebook utilization and prevent codebook collapse, whereas BAR does not appear to require such a loss.
Could you clarify the main differences between the FSQ formulation used in BAR and LFQ? In particular, are there any important architectural or optimization differences beyond the use of entropy regularization?
I am also curious about why entropy loss is not necessary in BAR. One possibility is that the encoder distillation or alignment objective with the frozen SigLIP2 encoder implicitly encourages diverse and informative code usage, thereby serving a role similar to entropy regularization. Another possibility is that entropy regularization becomes less important when the quantization dimensionality—and therefore the effective codebook size—is relatively modest, such as in the 16-bit setting.
I would greatly appreciate any clarification or insight from the authors. Thank you!