ByteDance open-sources BitDance to speed up AR image generation by 30x
ByteDance just open-sourced BitDance, an autoregressive image generation model that accelerates high-resolution synthesis by over 30x. By replacing codebook indices with binary visual tokens and decoding patches in parallel via continuous-space diffusion, it achieves a record 1.24 FID on ImageNet. Why it matters: AR models typically suffer from slow generation speeds, but BitDance proves they can rival diffusion in efficiency.
BitDance is an open-source autoregressive image generation model from ByteDance designed to fix the biggest problem in AR image synthesis: agonizingly slow inference speeds.
When generating 1024x1024 images, BitDance runs over 30x faster than traditional autoregressive generators. For developers who love AR's generation fidelity but cannot stomach the per-token latency, this is a massive leap forward.
How it delivers 30x faster generation
Standard AR models represent images using discrete codebook indices. If the codebook is too small, quality drops; if it is huge, predicting tokens via softmax becomes a computational nightmare.
BitDance bypasses codebooks entirely by mapping images into compact binary visual tokens. Thanks to high-entropy binary latents, each token can represent up to 2^256 states, capturing rich visual detail with minimal token footprint.
Sampling from such an enormous state space is impossible with standard classification heads. BitDance solves this by adding a binary diffusion head, replacing discrete classification with continuous-space diffusion to generate binary tokens.
On top of that, it introduces a decoding strategy called next-patch diffusion. Instead of decoding one visual token at a time, the model predicts 64 tokens in parallel during each diffusion step, slicing down overall latency.
Benchmarks and model specs
The numbers back up the design. On the ImageNet 256x256 benchmark, BitDance achieved an FID score of 1.24, setting a new record among autoregressive models.
Efficiency gains are equally dramatic. A compact 260M BitDance model beat state-of-the-art parallel AR baselines carrying 1.4B parameters, while using 5.4x fewer parameters and running 8.7x faster.
When scaling up to large multimodal training setups with models as large as 14B parameters, BitDance reliably generates photorealistic high-resolution visuals without losing its speed advantage. ByteDance has open-sourced both the codebase and model weights for the community.
Why it matters
We think BitDance is a smart architectural blueprint for the future of multimodal foundations. Autoregressive models naturally align with text LLMs, but diffusion has dominated image generation simply because AR was too slow to deploy economically. By fusing continuous diffusion heads directly into a binary AR tokenizer, BitDance removes the speed tax. AI engineers looking to build scalable text-to-image or unified multimodal pipelines should definitely give this repository a test run.
Sources (1)
- arxiv.org(opens in a new tab)Originalarxiv.org