DiverseAR
Boosting Diversity in Bitwise Autoregressive Image Generation
1 PRLab, Nanjing University ·
2 The University of Hong Kong ·
3 UCAS ·
4 Lovart AI ·
5 WeChat, Tencent Inc.
Abstract
In this paper, we investigate the underexplored challenge of sample diversity in
autoregressive (AR) generative models with bitwise visual tokenizers.
We initially analyze the factors limiting diversity in bitwise AR models
and identify two key issues:(1) the binary classification nature of
bitwise modeling, which restricts the prediction space, and (2)
the overly-sharp logits distribution, which causes sampling collapse and reduces diversity.
Built on these insights, we propose DiverseAR, a principle and effective method
that enhances image diversity without sacrificing visual quality. Specifically,
we introduce an adaptive logits distribution scaling mechanism that dynamically
adjusts the sharpness of the binary output distribution across different sampling steps,
resulting in a smoother prediction distribution and improved diversity. To mitigate
the potential fidelity loss caused by distribution smoothing, we further develop an
energy-based generation path search algorithm that avoids sampling low-confidence tokens,
thereby preserving high visual quality. Extensive experiments
highlight that DiverseAR can unlock greater diversity in bitwise autoregressive image generation.
Why Bitwise AR Model Degrades Diversity?
In bitwise autoregressive models, the predicted probabilities often become
overly peaked, giving one class near-certain confidence.
This causes top-p sampling to collapse into a deterministic choice, removing randomness.
As shown in the two subplots above, this collapse of bit-level randomness
restricts feature variation, and diversity degradation mainly arises from
(1) the binary classification nature of bitwise modeling and
(2) the overconfident output distributions.
* The diversity issue mainly originates from bitwise tokenization, rather than from VAR (Visual Autoregressive Modeling) itself.
DiverseAR Framework
DiverseAR is an effective approach that enhances image and video diversity
in bitwise autoregressive modeling without sacrificing visual quality.
It introduces adaptive logits smoothing and an energy-based generation path selection
strategy to achieve richer, more diverse sampling.
Qualitative Comparison (Image Generation)
Qualitative comparisons between Infinity generation and
DiverseAR show that integrating DiverseAR leads to a clear
improvement in diversity while preserving high visual quality.
Qualitative Comparison (Video Generation)
Vanilla (InfinityStar)
DiverseAR
Vanilla (InfinityStar)
DiverseAR
Vanilla (InfinityStar)
DiverseAR
Vanilla (InfinityStar)
DiverseAR
Qualitative comparisons in video generation show that
DiverseAR consistently improves motion and content diversity
while maintaining the visual fidelity of InfinityStar.
@article{yang2025diversear,
title={DiverseAR: Boosting Diversity in Bitwise Autoregressive Image Generation},
author={Yang, Ying and Lv, Zhengyao and Pan, Tianlin and Wang, Haofan and Yang, Binxin and Yin, Hubery and Li, Chen and Si, Chenyang},
journal={arXiv preprint arXiv:2512.02931},
year={2025}
}
.png)