Quantization Error Unreliable for Discrete Visual Token Semantic Loss

Research suggests that quantization error is an unreliable proxy for semantic loss in discrete visual representations. Downstream multimodal understanding performance is instead determined by the effective bit budget of discrete indices rather than raw reconstruction error.