Mirai: Autoregressive Visual Generation Needs Foresight

TL;DR

This work identifies a key limitation of autoregressive visual generation: purely causal next-token supervision impedes global coherence and slows convergence. It introduces Mirai, a foresight-enabled framework that aligns AR internal representations with future-aware signals in a 2D token grid, using two instantiations: Mirai-E (explicit foresight from an online EMA) and Mirai-I (implicit foresight from a frozen bidirectional encoder). Through extensive experiments on ImageNet with LlamaGen variants, Mirai markedly accelerates convergence (up to ~10x faster) and improves generation quality (FID reductions from ~5.34 to as low as 4.34 on 256×256), while maintaining same inference cost. The findings suggest that visual AR models benefit from foresight during training, leading to stronger global structure and practical gains in image fidelity and training efficiency.

Abstract

Autoregressive (AR) visual generators model images as sequences of discrete tokens and are trained with next token likelihood. This strict causality supervision optimizes each step only by its immediate next token, which diminishes global coherence and slows convergence. We ask whether foresight, training signals that originate from later tokens, can help AR visual generation. We conduct a series of controlled diagnostics along the injection level, foresight layout, and foresight source axes, unveiling a key insight: aligning foresight to AR models' internal representation on the 2D image grids improves causality modeling. We formulate this insight with Mirai (meaning "future" in Japanese), a general framework that injects future information into AR training with no architecture change and no extra inference overhead: Mirai-E uses explicit foresight from multiple future positions of unidirectional representations, whereas Mirai-I leverages implicit foresight from matched bidirectional representations. Extensive experiments show that Mirai significantly accelerates convergence and improves generation quality. For instance, Mirai can speed up LlamaGen-B's convergence by up to 10$\times$ and reduce the generation FID from 5.34 to 4.34 on the ImageNet class-condition image generation benchmark. Our study highlights that visual autoregressive models need foresight.

Figures (15)

• Figure 1: Left: The sample comparison between the AR baseline LlamaGen-B sun2024autoregressive and our Mirai with same 300-epoch training . The area enclosed by the red rectangle demonstrates the global consistency of images generated by our method. For example, in the rocket launch scene (bottom row right), the baseline model fails to maintain global structure, rendering a misaligned smoke. In contrast, our method generates a complete and structurally coherent result. Right: The performance of our Mirai on training acceleration.
• Figure 2: Overview of our explorations in the visual AR with foresight. For illustration, all subfigures (except the right of (c)) use $K = 3$ foresight tokens here. (a) Foresight injection level. (b) Foresight in 1D scan vs. 2D grid. (c) The source of foresight.
• Figure 3: Internal representation alignment with implicit foresight from bidirectional encoder. All experiments are performed on ImageNet 256×256 with a 50k-step training.
• Figure 4: Foresight token number analysis. All models are LlamaGen-B trained for 80 epochs.
• Figure 5: Two EMA selection strategies. All models are LlamaGen-B trained for 300 epochs.