Towards Scalable Pre-training of Visual Tokenizers

Focused on the VTP (Visual Tokenizer) backbone for high-fidelity generation.