Training-free method enables continuous bitrate control for scalable image coding

Researchers Yui Tatsumi and Hiroshi Watanabe from the University of Tokyo have introduced a training-free framework for continuous bitrate control in scalable image coding. The method, detailed in arXiv:2606.00158, addresses a key gap in variable-rate compression for systems that serve both human viewers and machine vision algorithms. Traditional scalable coding methods require extensive retraining to adjust bitrates, but this new approach dynamically alters quantization steps based on predicted scale values from the encoder. This allows the framework to smoothly vary compression rates without any additional training or fine-tuning. The system preserves high-scale (detail-rich) information in two distinct layers: a machine layer optimized for computer vision tasks (like object detection or segmentation) and an enhancement layer that improves visual quality for human viewing.

Experimental results validate the approach across several standard datasets, showing that the training-free method achieves competitive rate-distortion performance comparable to trained variable-rate models. The researchers also demonstrate the critical role of bitrate allocation between the two layers—improper allocation degrades performance for either humans or machines. By enabling continuous control (not just discrete presets), the framework is well-suited for bandwidth-constrained environments like edge devices, IoT cameras, or real-time video streaming where conditions change dynamically. This work could lead to more efficient compression pipelines that automatically adapt to both network conditions and the dual requirements of human and machine consumption.