aiqbench analysis
This research paper introduces SelfBootTok, a novel tokenization method that decomposes visual information into global and local groups to reduce computational redundancy in image generation tasks. By shifting the burden of local detail prediction to the tokenizer, the approach achieves a 40% reduction in computation while setting a new state-of-the-art gFID score with significantly fewer tokens. The work addresses fundamental efficiency challenges in diffusion models and generative AI architectures through improved representation learning strategies.
Balancing Image Compression and Generation with Bootstrapped Tokenization
arXiv:2606.05552v2 Announce Type: replace Abstract: Despite progress in image tokenization, standard methods encode redundant information by mixing all granularities within each token, thus redundancy persists between tokens. The mix of information of different granularity also complicates the training of generators. This paper introduces SelfBootTok, a method that resolves this by cleanly decomposing information into global and local token groups. Through self-bootstrapped learning, the model predicts local details exclusively from global tokens, shifting the burden of visual details from the generator to the tokenizer. Consequently, our generator is far more efficient, requiring only global tokens and reducing computation by approximately 40%, while delivering superior reconstruction and generation. Moreover, this paradigm scales elegantly: by leveraging more data or parameters to self-supervise local representation learning, SelfBootTok achieves a new state-of-the-art gFID score of 1.56 using only 64 tokens.
Abstract:Despite progress in image tokenization, standard methods encode redundant information by mixing all granularities within each token, thus redundancy persists between tokens. The mix of information of different granularity also complicates the training of generators. This paper introduces SelfBootTok, a method that resolves this by cleanly decomposing information into global and local token groups. Through self-bootstrapped learning, the model predicts local details exclusively from global tokens, shifting the burden of visual details from the generator to the tokenizer. Consequently, our generator is far more efficient, requiring only global tokens and reducing computation by approximately 40%, while delivering superior reconstruction and generation. Moreover, this paradigm scales elegantly: by leveraging more data or parameters to self-supervise local representation learning, SelfBootTok achieves a new state-of-the-art gFID score of 1.56 using only 64 tokens.
Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Graphics (cs.GR)
arXiv:2606.05552 [cs.LG]
(or arXiv:2606.05552v2 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2606.05552 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 227 words
Related coverage
Balancing Image Compression and Generation with Bootstrapped Tokenization
arXiv · Jun 6, 2026
Static Bootstrap Placement for Encrypted Language Model Decoding
arXiv · Oct 6, 2026
LightMTP: Lightweight Latent Multi-Token Prediction
arXiv · Oct 7, 2026
Directional Contextual Representations for Dependency Relations: Why Cross-Direction Pairing Fails
arXiv · Aug 24, 2026