reAR

Rethinking Visual Autoregressive Models via Generator–Tokenizer Consistency Regularization

ICLR 2026

Paper · Project page · Code

reAR addresses generator–tokenizer inconsistency: an autoregressive model can produce discrete token sequences that its visual tokenizer does not decode well. The method adds token-wise consistency regularization during training so the generator remains robust to imperfect contexts and learns representations that are compatible with the tokenizer embedding space.

It changes neither the tokenizer nor the generation order or inference path. This repository publishes the three released ImageNet checkpoints.

Checkpoints

Variant	Weight file	Role
reAR-S	reAR-S.safetensors	Small configuration
reAR-B	reAR-B.safetensors	Base configuration
reAR-L	reAR-L.safetensors	Large configuration

Model definitions, tokenizer setup, configuration files, sampling, and ADM evaluation live in the official codebase. The safetensors files are weights rather than self-contained Transformers packages.

Method

During training, reAR combines the ordinary next-token objective with two compatible regularizers:

• Noisy-context regularization exposes the causal generator to perturbed histories while preserving parallel training.
• Codebook-embedding regularization guides intermediate representations toward the tokenizer's visual embedding space.

The auxiliary training branches are removed at inference, leaving the standard causal token-generation and tokenizer-decoding path shown on the right.

Quick start

Install and use the official implementation, then download the checkpoint that matches the selected configuration:

git clone https://github.com/QY-H00/reAR.git
cd reAR

conda create -n rear python=3.10
conda activate rear
pip install -r requirements.txt

hf download qyoo/reAR reAR-L.safetensors --local-dir ckpt/reAR

Follow the repository's scripts_bash/rear_test.sh workflow for sampling and evaluation. The MaskGiT-VQGAN tokenizer checkpoint required by the reference configuration is downloaded separately, as documented in the code repository.

Paper results

On 256 × 256 class-conditional ImageNet generation, the paper reports that reAR reduces gFID from 3.02 to 1.86 and raises IS to 316.9 with a standard rasterization-based tokenizer. With an advanced tokenizer, the reported gFID is 1.42 using a 177M-parameter generator. These are paper-level results; use the matching code, tokenizer, model configuration, and evaluation protocol when reproducing them.

Intended use and limitations

The released checkpoints are intended for research on class-conditional visual autoregressive generation and generator–tokenizer consistency. They are not general text-to-image models. Their behavior inherits limitations and biases from ImageNet, the selected visual tokenizer, autoregressive sampling, and the evaluation pipeline. Generated images may contain artifacts or misleading content and should be reviewed before downstream use.

Citation

@article{he2025rear,
  title   = {REAR: Rethinking Visual Autoregressive Models via Generator-Tokenizer Consistency Regularization},
  author  = {He, Qiyuan and Li, Yicong and Ye, Haotian and Wang, Jinghao and Liao, Xinyao and Heng, Pheng-Ann and Ermon, Stefano and Zou, James and Yao, Angela},
  journal = {arXiv preprint arXiv:2510.04450},
  year    = {2025}
}

License

Apache 2.0. Review the official repository for dependency licenses and third-party acknowledgements.