VIRAL: 通过类比在扩散变换器中实现视觉上下文推理
VIRAL: Visual In-Context Reasoning via Analogy in Diffusion Transformers
AI总结：
VIRAL通过类比和扩散变换器实现视觉上下文推理，解决视觉任务异质性问题，提升开放域编辑性能。
摘要：
在计算机视觉中复制上下文学习（ICL）仍然具有挑战性，因为任务异质性很高。我们提出了VIRAL框架，通过将ICL作为条件生成来激发预训练图像编辑模型中的视觉推理（x_s : x_t :: x_q : y_q）。我们通过角色感知的多图像条件化适配冻结的扩散变换器（DiT），并引入混合专家LoRA来减轻跨多样任务的梯度干扰。此外，为了弥补当前视觉上下文数据集的差距，我们编纂了一个大规模数据集，涵盖感知、修复和编辑。实验表明，VIRAL优于现有方法，验证了统一的V-ICL范式可以处理大多数视觉任务，包括开放域编辑。我们的代码可在https://anonymous.4open.science/r/VIRAL-744A获取。
英文摘要：
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
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
