Token Translation, Stage by Stage
Step through the three stages. Blue blocks are trained at that stage, grey ones are frozen. Drag the dot in the lattice to feel what FSQ does.
FSQ lattice: continuous z → nearest lattice point
Finite scalar quantization rounds each latent channel to one of a few fixed levels. There is no learned codebook. The frozen target decoder maps the predicted tokens into the target modality. The tokenizer's own reconstruction ceiling is 0.214 LPIPS, far below any method's translation error: representation is not the bottleneck, prediction is.