We present Visual AutoRegressive modeling (VAR), a new generation paradigm that redefines the autoregressive learning on images as coarse-to-fine "next-scale prediction" or "next-resolution prediction", diverging from the standard raster-scan "next-token prediction". This simple, intuitive methodology allows autoregressive (AR) transformers to learn visual distributions fast and generalize well: VAR, for the first time, makes GPT-like AR models surpass diffusion transformers in image generation. On ImageNet 256x256 benchmark, VAR significantly improve AR baseline by improving Frechet inception distance (FID) from 18.65 to 1.73, inception score (IS) from 80.4 to 350.2, with around 20x faster inference speed. It is also empirically verified that VAR outperforms the Diffusion Transformer (DiT) in multiple dimensions including image quality, inference speed, data efficiency, and scalability. Scaling up VAR models exhibits clear power-law scaling laws similar to those observed in LLMs, with linear correlation coefficients near -0.998 as solid evidence. VAR further showcases zero-shot generalization ability in downstream tasks including image in-painting, out-painting, and editing. These results suggest VAR has initially emulated the two important properties of LLMs: Scaling Laws and zero-shot task generalization. We have released all models and codes to promote the exploration of AR/VAR models for visual generation and unified learning.
核心贡献 · Key contributions
提出视觉自回归（VAR）建模，采用下一尺度预测，将图像上的自回归学习重新定义为从粗到细的生成。 Proposes Visual AutoRegressive (VAR) modeling with next-scale prediction, redefining autoregressive learning on images as coarse-to-fine generation.
在 ImageNet 256×256 上实现了最先进的图像生成，FID 为 1.73，IS 为 350.2，首次超越了扩散 Transformer。 Achieves state-of-the-art image generation on ImageNet 256×256 with FID 1.73 and IS 350.2, surpassing diffusion transformers for the first time.
展示了 VAR 模型清晰的幂律缩放定律，线性相关系数接近-0.998，类似于大语言模型。 Demonstrates clear power-law scaling laws for VAR models, with linear correlation coefficients near -0.998, similar to LLMs.
在下游任务如图像修复、外推和编辑中展示了零样本泛化能力，无需任务特定调优。 Shows zero-shot generalization in downstream tasks like image in-painting, out-painting, and editing without task-specific tuning.
与标准自回归模型相比，推理速度提升 20 倍，计算复杂度从 O(n^6)降至 O(n^4)。 Provides 20x faster inference speed compared to standard autoregressive models, reducing computational complexity from O(n^6) to O(n^4).
发布所有模型和代码，以促进 AR/VAR 模型在视觉生成和统一学习中的探索。 Releases all models and code to promote exploration of AR/VAR models for visual generation and unified learning.
局限 · Limitations
VAR 依赖于固定的多尺度分词方案；性能可能因不同的分词器设计而异。 VAR relies on a fixed multi-scale tokenization scheme; performance may vary with different tokenizer designs.
研究集中于类别条件图像生成；文本提示生成尚未探索。 The study focuses on class-conditional image generation; text-prompt generation is not yet explored.
视频生成尚未实现；扩展到视频需要进一步研究 3D 下一尺度预测。 Video generation is not implemented; extension to video requires further research on 3D next-scale prediction.
VQVAE 分词器保持与基线相同；先进的分词器可能进一步提升性能。 The VQVAE tokenizer is kept unchanged from baseline; advanced tokenizers could further improve performance.
缩放定律仅在 ImageNet 上验证；对其他数据集或领域的泛化性未验证。 Scaling laws are validated only on ImageNet; generalization to other datasets or domains is not verified.
论文章节 · Sections（共 16）
摘要Abstract
1 引言1 Introduction
2.1 大型自回归语言模型的性质2.1 Properties of large autoregressive language models
2.2 视觉生成2.2 Visual generation
3.1 预备知识：基于下一词元预测的自回归建模3.1 Preliminary: autoregressive modeling via next-token prediction
3.2 视觉自回归建模：基于下一尺度预测3.2 Visual autoregressive modeling via next-scale prediction
3.3 实现细节3.3 Implementation details
4 实验结果4 Empirical Results
4.1 最先进的图像生成4.1 State-of-the-art image generation
4.2 幂律缩放定律4.2 Power-law scaling laws
4.3 零样本任务泛化4.3 Zero-shot task generalization
4.4 消融研究4.4 Ablation Study
5 未来工作5 Future Work
6 结论6 Conclusion
附录 A VQVAE 中的词元依赖Appendix A Token dependency in VQVAE
附录 B AR 与 VAR 生成的时间复杂度Appendix B Time complexity of AR and VAR generation