A new preprint reports favorable results from adding adapter modules to a frozen image-compression system for machine-vision tasks. On the paper's main TIC benchmark, its Ours-Base configuration reported a -92.788% BD-rate and 17.575 BD-acc for classification, a -65.607% BD-rate and 4.249 BD-mAP for detection, and a -62.589% BD-rate and 3.624 BD-mAP for segmentation, with 0.32 million trainable parameters.
The document is an arXiv v1 preprint dated 26 August 2026. The supplied analysis reports no confidence intervals or repeated-run variability for these computational benchmark results.
A smaller training footprint
The proposed system, CrossMambaTuning, targets deployment and retraining overhead in adapting pretrained learned image-compression codecs. It combines a task-aware Mamba adapter with a Scale-Invariant Cross-Layer Adapter, or SICA, for cross-layer fusion.
The authors reported a 72% reduction in parameter overhead versus the cited state-of-the-art method for the Tiny variant, which used 0.08 million trainable parameters. A separate comparison lists 0.08 million trainable parameters for Tiny against 7.51 million for full fine-tuning.
Training froze the pretrained codec and optimized only the inserted adapters under a rate-distortion objective. The evaluation used ImageNet for image classification and COCO2017 for object detection and instance segmentation.
Results across two codec designs
The comparison used a simplified Lu2022-TIC hyperprior codec and CNN-based ELIC with spatial-channel context, alongside full fine-tuning and named parameter-efficient baseline methods.
The reported results also carried over to the ELIC object-detection benchmark. Ours-Tiny reported a -57.495% BD-rate, 3.863 BD-mAP and 0.11 million trainable parameters. Ours-Base reported a -62.710% BD-rate, 4.500 BD-mAP and 0.43 million trainable parameters.
The paper included feature-distribution diagnostics for the APC component. With APC, Wasserstein values changed from 0.0308 to 0.0277 for intermediate stages and from 0.0388 to 0.0318 for the latent representation. KL-divergence values changed from 0.0396 to 0.0318 for stages and from 0.0163 to 0.0077 for the latent representation. These are distribution-alignment diagnostics rather than direct task outcomes.
What the component tests showed
The all-component ablation row containing APC, LIE, SICA and TSPG was reported as the strongest core configuration. It recorded -65.607% detection BD-rate and 4.249 BD-mAP, along with -62.589% segmentation BD-rate and 3.624 BD-mAP, using 0.32 million trainable parameters.
In the parameter-sharing test, shared SICA parameters reported -65.607% BD-rate, 4.249 BD-mAP and 0.32 million parameters. Independent modules reported -63.224% BD-rate, 4.106 BD-mAP and 0.36 million parameters.
A separate spatial-correlation analysis, used as a proxy for redundancy, reported values of 0.18 for the previous state-of-the-art method, 0.05 for the proposed approach and 0.01 for full fine-tuning. The paper treats this as an indirect diagnostic rather than a direct measure of deployment or task benefit.
The trade-off behind the headline numbers
The reconstruction figures are presented as references rather than a full rate-distortion comparison. On the ImageNet-val reference, the base codec and the adapter-removed setting both reported 0.227 bpp, 28.92 dB PSNR and 0.971 MS-SSIM. With the adapters attached, the reported values were 0.206 bpp, 23.90 dB and 0.923.
That result sets an important boundary around the benchmark claims: the supplied analysis does not show that human-view reconstruction quality is preserved when the adapters remain attached. The reported evidence is limited to the named ImageNet and COCO2017 tasks and the TIC and ELIC codec comparisons.
The training schedules used Adam with batch sizes of 16 for classification and 8 for detection and segmentation. The corresponding schedules ran for 8, 40 and 40 epochs, with a base learning rate of 4e-4 for all three tasks.
Taken together, the study presents parameter efficiency and favorable task metrics as the main advantages of its adapter design, while leaving uncertainty, reconstruction trade-offs and performance outside the tested settings unresolved.
The work reports support from the China Postdoctoral Science Foundation under Grant 2024M752531 and the National Natural Science Foundation of China under Grant 62171342.
Paper data and sources
Original title: CrossMambaTuning: Synergistic Spatial and Cross-Layer Adaptation for Machine Vision Compression
Authors: Haobo Xiong, Shaobo Liu, Kai Liu, Chongyang Ding
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text