VIRAL: Visual In-Context Reasoning via Analogy in Diffusion Transformers

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Li, Zhiwen, Duan, Zhongjie, Ye, Jinyan, Chen, Cen, Chen, Daoyuan, Li, Yaliang, Chen, Yingda
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910010028589056
author Li, Zhiwen
Duan, Zhongjie
Ye, Jinyan
Chen, Cen
Chen, Daoyuan
Li, Yaliang
Chen, Yingda
author_facet Li, Zhiwen
Duan, Zhongjie
Ye, Jinyan
Chen, Cen
Chen, Daoyuan
Li, Yaliang
Chen, Yingda
contents Replicating In-Context Learning (ICL) in computer vision remains challenging due to task heterogeneity. We propose \textbf{VIRAL}, a framework that elicits visual reasoning from a pre-trained image editing model by formulating ICL as conditional generation via visual analogy ($x_s : x_t :: x_q : y_q$). We adapt a frozen Diffusion Transformer (DiT) using role-aware multi-image conditioning and introduce a Mixture-of-Experts LoRA to mitigate gradient interference across diverse tasks. Additionally, to bridge the gaps in current visual context datasets, we curate a large-scale dataset spanning perception, restoration, and editing. Experiments demonstrate that VIRAL outperforms existing methods, validating that a unified V-ICL paradigm can handle the majority of visual tasks, including open-domain editing. Our code is available at https://anonymous.4open.science/r/VIRAL-744A
format Preprint
id arxiv_https___arxiv_org_abs_2602_03210
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle VIRAL: Visual In-Context Reasoning via Analogy in Diffusion Transformers
Li, Zhiwen
Duan, Zhongjie
Ye, Jinyan
Chen, Cen
Chen, Daoyuan
Li, Yaliang
Chen, Yingda
Computer Vision and Pattern Recognition
Replicating In-Context Learning (ICL) in computer vision remains challenging due to task heterogeneity. We propose \textbf{VIRAL}, a framework that elicits visual reasoning from a pre-trained image editing model by formulating ICL as conditional generation via visual analogy ($x_s : x_t :: x_q : y_q$). We adapt a frozen Diffusion Transformer (DiT) using role-aware multi-image conditioning and introduce a Mixture-of-Experts LoRA to mitigate gradient interference across diverse tasks. Additionally, to bridge the gaps in current visual context datasets, we curate a large-scale dataset spanning perception, restoration, and editing. Experiments demonstrate that VIRAL outperforms existing methods, validating that a unified V-ICL paradigm can handle the majority of visual tasks, including open-domain editing. Our code is available at https://anonymous.4open.science/r/VIRAL-744A
title VIRAL: Visual In-Context Reasoning via Analogy in Diffusion Transformers
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2602.03210