Image denoising has made steady progress with deep learning, but textures remain a stubborn edge case. A new paper on arXiv argues that the fine patterns and repetitive structures found in textures are especially vulnerable to over-smoothing, even in state-of-the-art denoisers. The authors note that while the human brain reconstructs such details effortlessly, algorithms often wash them out.

The researchers evaluated current denoisers on texture datasets and found that meaningful performance gains came only after fine-tuning on those datasets. However, that fine-tuning came at a cost: the models catastrophically forgot their original training data. This trade-off, they argue, means textures cannot be treated as just another denoising benchmark.

Because denoising quality directly affects downstream tasks such as generative modeling, the paper makes a case for building dedicated texture-aware denoisers rather than relying on generic ones. The study highlights a gap in current denoising research and suggests that specialized architectures or training strategies may be needed for faithful texture reconstruction and synthesis.