vqrae-representation-quantization

Unify multimodal understanding, generation, and reconstruction using high-dimensional codebooks for semantic information. VQRAE achieves 100% codebook utilization at 1536 dimensions—ideal when you need a single tokenizer for vision-language tasks.

Independent GitHub snapshot · Automated classification · Original documentation governs