CAST-LUT: Tokenizer-Guided HSV Look-Up Tables for Purple Flare Removal

TL;DR

This work tackles purple flare artifacts that degrade image color fidelity around highlights, a problem exacerbated by color coupling and data scarcity. It introduces CAST-LUT, a two-stage framework that uses Chroma-Aware Spectral Tokenizer (CAST) to produce discrete semantic tokens from the Hue and Value channels, which then guide a set of decoupled 1D-LUTs operating in HSV space for targeted color correction. The authors contribute the PFSD dataset, a suite of artifact- and color-focused metrics (including PSNR-F, PSNR-NF, and Hue Alignment Error), and a loss that explicitly emphasizes purple flare suppression. Empirically, CAST-LUT achieves state-of-the-art performance on PFSD across standard and artifact-specific metrics, while enabling real-time mobile deployment, highlighting practical impact for camera pipelines.

Abstract

Purple flare, a diffuse chromatic aberration artifact commonly found around highlight areas, severely degrades the tone transition and color of the image. Existing traditional methods are based on hand-crafted features, which lack flexibility and rely entirely on fixed priors, while the scarcity of paired training data critically hampers deep learning. To address this issue, we propose a novel network built upon decoupled HSV Look-Up Tables (LUTs). The method aims to simplify color correction by adjusting the Hue (H), Saturation (S), and Value (V) components independently. This approach resolves the inherent color coupling problems in traditional methods. Our model adopts a two-stage architecture: First, a Chroma-Aware Spectral Tokenizer (CAST) converts the input image from RGB space to HSV space and independently encodes the Hue (H) and Value (V) channels into a set of semantic tokens describing the Purple flare status; second, the HSV-LUT module takes these tokens as input and dynamically generates independent correction curves (1D-LUTs) for the three channels H, S, and V. To effectively train and validate our model, we built the first large-scale purple flare dataset with diverse scenes. We also proposed new metrics and a loss function specifically designed for this task. Extensive experiments demonstrate that our model not only significantly outperforms existing methods in visual effects but also achieves state-of-the-art performance on all quantitative metrics.

Figures (16)

• Figure 1: Color space analysis and CAST-guided correction comparison. (a) RGB vs. HSV: HSV decouples purple flare characteristics more clearly. (b) CAST-guided effect: The CAST module significantly improves performance.
• Figure 2: Overall architecture of the decoupled CAST-LUT network. The CAST module first extracts semantic tokens from the input image's H and V channels via an Encoder-VQ pipeline. These tokens then guide the dynamic generation and weighted fusion of multiple, decoupled 1D-LUTs for the H, S, and V channels to obtain the final corrected image.
• Figure 3: The generation process of the penalty mask $\mathbf{M}$ used in our flare suppression loss $\mathcal{L}_f$. The final mask (d) is obtained by element-wise multiplication of the edge map (b) and the purple weight map (c), which can precisely locate flare artifacts and apply targeted loss.
• Figure 4: Qualitative comparison with state-of-the-art methods on our PFSD dataset.
• Figure 5: Visualization of learned semantic tokens from the CAST module.