Pinned post
Discrete or continuous tokens? Or even tokenizer-free? The visual modeling debate rages on, but for now, let me introduce L24SQ, a provably optimal, regularizer-free quantizer with a large codebook (~200k), achieving SoTA reconstruction-compression tradeoff and generative power!