DiverseAR: Boosting Diversity in Bitwise Autoregressive Image Generation

TL;DR

This paper addresses the limited diversity in bitwise autoregressive image models by analyzing key issues and proposing DiverseAR, which adaptively smooths logits and uses energy-based search to enhance diversity without losing quality.

Contribution

The paper introduces DiverseAR, a novel method that improves sample diversity in bitwise AR models through adaptive logits scaling and energy-based path search.

Findings

Significant increase in image diversity with maintained quality

Adaptive logits scaling effectively smooths prediction distributions

Energy-based search preserves high-fidelity image generation

Abstract

In this paper, we investigate the underexplored challenge of sample diversity in autoregressive (AR) generative models with bitwise visual tokenizers. We first analyze the factors that limit 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. Building on these insights, we propose DiverseAR, a principled 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 during sampling, resulting in smoother predictions and greater diversity. To mitigate 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 demonstrate that DiverseAR substantially improves sample diversity in bitwise autoregressive image generation.