Towards Scalable Pre-training of Visual Tokenizers

paper
MiniMax
Focused on the VTP (Visual Tokenizer) backbone for high-fidelity generation.
Paper (arXiv)GitHub
visionarchitectureresearch