Logit Refiner: Improving Visual Autoregressive Models via Intra-Scale Dependency Modeling

MCML Authors

Stefan Baumann

→ Group Björn Ommer
 Computer Vision & Learning

Felix Krause

→ Group Björn Ommer
 Computer Vision & Learning

Björn Ommer

Prof. Dr.

Core PI

Computer Vision & Learning

Abstract

Visual Autoregressive Models (VAR) generate images through next-scale prediction, producing all tokens within each scale in parallel. We show that this parallel decoding constitutes a mean-field-style approximation that discards spatial dependencies among same-scale tokens, causing locally incoherent samples regardless of backbone capacity – a limitation of the decoding rule. Addressing this limitation, we introduce the Logit Refiner, a lightweight autoregressive module that restores intra-scale dependencies by sequentially sampling tokens conditioned on frozen backbone features. Adding only 10% parameters and less than 5% of the base model’s training compute, it plugs into any pretrained VAR checkpoint without retraining. Controlled ablations isolate joint intra-scale sampling – rather than additional capacity or training – as the critical ingredient. Across backbones from 310M to 2B parameters on class-conditional ImageNet 256 x 256, the refiner consistently improves generation quality, enabling a 1.1B-parameter model to surpass one twice its size. The approach further generalizes to text-to-image generation, confirming that the mean-field bottleneck persists across VAR variants and is effectively alleviated by our method.

inproceedings LBK+26

---

ECCV 2026

19th European Conference on Computer Vision. Malmö, Sweden, Sep 08-12, 2026.

---

Authors

M. Li • S. A. Baumann • F. Krause • B. Ommer

---

Links

DOI

---

Research Area

 B1 | Computer Vision