Variance-normalized latent distillation (VNLD) for domain-specific learned image compression under JPEG AI constraints

We introduce variance-normalized latent distillation (VNLD), a new loss function for learned image compression. VNLD transfers rich latent representations from a teacher to a student through channelwise variance-normalized alignment, which amplifies informative channels and stabilizes training, while keeping the decoder frozen to comply with JPEG AI decoding requirements. The teacher is obtained by fine-tuning SegPIC, a state-of-the-art compression model, from its best pretrained checkpoint on one image category with high distortion weight (not rate–distortion (RD) constrained), and the student is initialized from the same checkpoint and optimized with the combined RD and VNLD losses. Experiments on the three dominant domains of smartphone photography (selfies, food, landscapes) show that VNLD achieves up to 6.77% BD-rate savings over the pretrained baseline and outperforms standard fine-tuning, while preserving performance on general benchmarks (Kodak, JPEG AI) and remaining robust on unseen datasets

within the same category. These results position VNLD as a candidate loss for future JPEG AI-compliant learned image compression.