ICLR 2026

Autoregressive Image Generation with Randomized Parallel Decoding

One causal architecture for random-order modeling, parallel decoding, and zero-shot spatial control.

Haopeng Li Jinyue Yang Guoqi Li Huan Wang
Westlake University · Institute of Automation, Chinese Academy of Sciences
Paper Code Models
ARPG generation samples, efficiency comparison, controllable generation, inpainting, outpainting, and resolution expansion results
Generation quality, efficiency, and zero-shot spatial control. ImageNet-1K · 256 × 256
1.83FID on ImageNet-1K
32parallel sampling steps
30×faster inference
−75%memory consumption

Abstract

Conventional visual autoregressive models inherit a fixed raster order. ARPG removes that constraint while retaining a causal architecture and its efficient KV cache.

ARPG is a visual autoregressive model for randomized parallel generation. Its decoupled decoding framework separates positional guidance from content representation, encoding them as queries and key-value pairs. Injecting that guidance directly into causal attention enables fully random-order training and generation without bidirectional attention.

The same design supports multiple simultaneous queries over a shared KV cache, turning random-order generation into efficient parallel inference. It also generalizes without task-specific training to image inpainting, outpainting, and resolution expansion.

How it works

ARPG treats “where to predict” and “what to predict” as two different signals, then uses that separation to unlock random-order and parallel decoding.

01

Decouple position from content

Next-token positions are encoded as queries, while visible image content stays in key-value representations.

02

Train in randomized order

Explicit positional guidance lets a causal transformer learn arbitrary token orders without switching to bidirectional attention.

03

Decode many tokens at once

Multiple position queries share one KV cache, enabling parallel prediction with lower latency and memory overhead.

One model, many layouts

Randomized order is more than a speed optimization: it makes generation naturally compatible with arbitrary known and unknown image regions.

G

Class-conditional generation

High-quality ImageNet synthesis through efficient parallel autoregressive decoding.

E

Image editing

Preserve visible tokens and predict new content in selected regions.

I

Inpainting & outpainting

Fill interior masks or extend beyond the original canvas without task-specific retraining.

R

Resolution expansion

Reuse the same random-order mechanism to add detail beyond the training resolution.

Model zoo

Pretrained checkpoints for ImageNet-1K at 256 × 256 are available on Hugging Face.

ModelParametersScheduleStepsCheckpoint
ARPG-L320MCosine64Download ↗
ARPG-XL719MCosine64Download ↗
ARPG-XXL1.3BCosine64Download ↗

Quick start

Load ARPG through the Hugging Face Diffusers pipeline and generate a batch of ImageNet classes in a few lines.

Full instructions 
from diffusers import DiffusionPipeline

pipeline = DiffusionPipeline.from_pretrained(
    "hp-l33/ARPG",
    custom_pipeline="hp-l33/ARPG"
)

image = pipeline(
    model_type="ARPG-XL",
    seed=0,
    num_steps=64,
    class_labels=[207, 360, 388, 113],
    cfg_scale=4,
    sample_schedule="arccos"
)

image.show()

Citation

If ARPG is useful in your research, please cite the ICLR 2026 paper.

@inproceedings{li2026autoregressive,
  title     = {Autoregressive Image Generation with Randomized Parallel Decoding},
  author    = {Haopeng Li and Jinyue Yang and Guoqi Li and Huan Wang},
  booktitle = {The Fourteenth International Conference on Learning Representations},
  year      = {2026}
}