ByteDouBao's new image Tokenizer: only 32 tokens are needed to generate an image, and the speed is increased by up to 410 times.

The AIxiv column is a column where this site publishes academic and technical content. In the past few years, the AIxiv column of this site has received more than 2,000 reports, covering top laboratories from major universities and companies around the world, effectively promoting academic exchanges and dissemination. If you have excellent work that you want to share, please feel free to contribute or contact us for reporting. Submission email: liyazhou@jiqizhixin.com; zhaoyunfeng@jiqizhixin.com

In the rapid development of generative models, Image Tokenization plays a very important role, such as VAE that Diffusion relies on or VQGAN that Transformer relies on. . These Tokenizers encode the image into a more compact latent space, making it more efficient to generate high-resolution images.

However, existing Tokenizers usually map the input image to a downsampled 2D matrix in the latent space. This design implicitly limits the mapping relationship between tokens and images, making it difficult to Effectively utilize redundant information in the image (for example, adjacent areas often have similar features) to obtain a more effective image encoding.

In order to solve this problem, ByteDance Beanbao Big Model Team and Technical University of Munich proposed a new 1D image Tokenizer: TiTok. This Tokenizer breaks the design limitations of 2D Tokenizer and can compress the entire image to a more compact Token sequence.

• Paper link: https://arxiv.org/abs/2406.07550
• Project link: https://yucornetto.github.io/projects/titok.html
• Code link: https://github.com/bytedance/1d-tokenizer

For a 256 x 256 resolution image, TiTok only needs a minimum of 32 Tokens to express it, which is 256 or 1024 Tokens than the usual 2D Tokenizer significantly reduced. For a 512 x 512 resolution image, TiTok requires at least 64 Tokens, which is 64 times smaller than Stable Diffusion’s VAE Tokenizer. In addition, on the task of ImageNet image generation, using TiTok as the Tokenizer generator has significantly improved the generation quality and generation speed.

At 256 resolution, TiTok achieved an FID of 1.97, significantly exceeding MaskGIT’s 4.21 using the same generator. At 512 resolution TiTok can achieve an FID of 2.74, which not only exceeds DiT (3.04), but also accelerates image generation by an astonishing 410 times compared to DiT! The best variant of TiTok achieved an FID of 2.13, significantly exceeding DiT while still achieving a 74x acceleration.

                                                                                                                                                        ^{​                                                                                                                                                                                                                                                                           with tokens required for images to result in significantly faster generation speeds , but while maintaining high-quality image generation.}

Model structure

The structure of TiTok is very simple. The encoder and decoder parts are each a ViT. During the encoding process, a set of latent tokens will be spliced ​​after the image patches. After passing through the encoder, only the latent tokens are retained and the quantization process is performed. The obtained quantized latent tokens will be spliced ​​together with a set of mask tokens and sent to the decoder to reconstruct the image from the mask token sequence.

Study on the properties of 1D Tokenization

The researchers conducted a series of experimental studies on different numbers of tokens used to represent images, different tokenizer sizes, reconstruction performance, generation performance, linear probing accuracy, and training and Comparison of reasoning speed. During this process, the researchers found that (1) only 32 Tokens can achieve good reconstruction and generation effects (2) By increasing the model size of Tokenizer, researchers can use fewer Tokens to represent images ( 3) When fewer Tokens are used to represent pictures, Tokenizer will learn stronger semantic information. (4) When fewer Tokens are used to represent pictures, training and inference speeds are significantly improved.

In addition, the video shows the reconstructed images using different Tokenizer sizes and the number of Tokens. It can be seen that a larger Tokenizer can reconstruct better quality images with limited Tokens. In addition, when there are only limited tokens, the model is more inclined to retain salient areas and achieve better reconstruction results.

Experimental verification

The researchers mainly compared with other methods at the 256 x 256 resolution and 512 x 512 resolution of ImageNet-1k. It can be seen that although TiTok uses a limited number of Tokens, it can achieve comparable reconstruction results (rFID) with other methods that use more Tokens. Using a smaller number of Tokens allows TiTok to maintain a higher generated image quality (gFID) At the same time, it has a significantly faster generation speed than other methods.

For example, TiTok-L-32 achieved a gFID score of 2.77 and can generate images at a speed of 101.6 images per second, which is significantly faster than other Diffusion Models (169 times faster than DiT) or Transformer Models (339 times faster than ViT-VQGAN).

TiTok’s advantage of using fewer tokens is more obvious in higher-resolution image generation, where TiTok-L-64 can be completed using only 64 tokens Reconstruction and generation of high-quality 512 resolution images. The quality of the generated images is not only higher than DiT (2.74 v.s. 3.04), but the generation speed is increased by nearly 410 times.

Conclusion

In this article, the researcher focuses on a new 1D Image Tokenizer and proposes a new Tokenizer to break the limitations of the existing 2D Tokenizer and make it more advanced Good use of redundant information in images. TiTok only needs a small number of Tokens (such as 32) to represent images, while still being able to perform high-quality image reconstruction and generation. In ImageNet's 256 resolution and 512 resolution generation experiments, TiTok not only achieved generation quality that exceeded Diffusion Models, but also achieved a hundred times faster generation speed.

About the Doubao Large Model Team

ByteDance Doubao Large Model Team was established in 2023 and is committed to developing the industry's most advanced AI large model technology and becoming a world-class research team. Contribute to technological and social development.

The Doubao Big Model team has long-term vision and determination in the field of AI. Its research directions cover NLP, CV, speech, etc., and it has laboratories and research positions in China, Singapore, the United States and other places. Relying on the platform's sufficient data, computing and other resources, the team continues to invest in related fields. It has launched a self-developed general large model to provide multi-modal capabilities. It supports 50+ businesses such as Doubao, Buttons, and Jimeng downstream, and is open to the public through the Volcano Engine. Corporate customers. At present, Doubao APP has become the AIGC application with the largest number of users in the Chinese market.

Welcome to join the Bytedance Beanbao Big Model Team, click the link below to enter the Bytedance Top Seed plan:

https://mp.weixin.qq.com/s/ZjQ-v6reZXhBP6G27cbmlQ

The above is the detailed content of ByteDouBao's new image Tokenizer: only 32 tokens are needed to generate an image, and the speed is increased by up to 410 times.. For more information, please follow other related articles on the PHP Chinese website!