shallowdream204/BitDance

What it solves

BitDance addresses common bottlenecks in discrete autoregressive (AR) models for visual generation, specifically poor tokenizer reconstruction, the difficulty of sampling from massive vocabularies, and slow generation speeds caused by token-by-token prediction.

How it works

BitDance uses a decoder-only architecture that combines three core innovations:

1. Large-Vocabulary Binary Tokenizer: A tokenizer that represents visual data using binary tokens to improve reconstruction quality.
2. Binary Diffusion Head: A specialized head that enables efficient sampling within a large discrete space.
3. Next-Patch Diffusion Paradigm: Instead of predicting one token at a time, this approach allows the model to predict multiple visual tokens in parallel (up to 64 per step), significantly accelerating the generation process.

Who it’s for

It is designed for researchers and developers working on multimodal generative models, specifically those looking for high-resolution, photorealistic image generation with faster inference speeds than standard autoregressive models.

Highlights

• High Speed: Achieves over 30x speedup compared to standard next-token prediction AR models.
• Parallel Prediction: Capable of predicting up to 64 visual tokens in a single step.
• Scalable Architecture: An open-source foundation model with 14B parameters trained on large-scale multimodal tokens.
• High Resolution: Supports the generation of photorealistic images up to 1024px.