LUTFormer: Lookup table transformer for image enhancement

초록

Existing image enhancement methods based on 3D lookup tables (LUTs) often yield suboptimal results by oversimplifying image context into a single global feature and disrupting the inherent geometric structure of a LUT during regression. To address these issues, we propose LUTFormer, a novel framework that reframes LUT prediction as a query-based refinement task. LUTFormer preserves geometric integrity by initializing LUT grid points as structured query tokens, which are then progressively refined by a transformer decoder. This decoder leverages a novel progressive cross-attention mechanism to inject multi-level image context, yielding a context-aware LUT transformation. Extensive experiments on multiple benchmark datasets confirm the effectiveness and efficiency of the proposed LUTFormer. The source code is available at https://github.com/Jinwon-Ko/LUTFormer.

키워드