Скачать или смотреть BitDance: Scaling Autoregressive Generative Models with Binary Tokens

The document introduces BitDance, a scalable autoregressive (AR) image generator designed to overcome limitations in visual AR modeling. BitDance addresses challenges like token design and inference efficiency by predicting high-entropy binary visual tokens, which allow each token to represent up to 2^256 states, yielding a compact yet highly expressive discrete representation. To handle sampling from such a vast token space, the model employs a novel binary diffusion head that uses continuous-space diffusion instead of standard classification. Furthermore, BitDance proposes "next-patch diffusion," an innovative decoding method enabling accurate parallel prediction of multiple tokens, which significantly accelerates inference. On ImageNet 256x256, BitDance achieves an FID of 1.24, setting a new benchmark for AR models. It demonstrates superior efficiency, outperforming state-of-the-art parallel AR models with 5.4 times fewer parameters and an 8.7 times speedup. For text-to-image generation, BitDance trains on large-scale multimodal tokens and generates high-resolution, photorealistic images efficiently. Specifically, it achieves over 30 times speedup compared to prior AR models when generating 1024x1024 images, showcasing strong performance and favorable scaling. The authors release code and models to facilitate further research.
#BitDance #AutoregressiveModels #ImageGeneration #BinaryTokens #DiffusionModels #DeepLearning #AI #Scalability #TextToImage

Donats:   / luxak  

paper - https://github.com/shallowdream204/Bi...
subscribe - https://t.me/arxivpaper
created with NotebookLM