Generative image communication transmits compact semantic tokens under a limited packet budget, and token selection directly determines the reconstruction quality after the full packet is decoded. Conventional approaches require repeated receiver‑side reconstructions for every candidate token to estimate its terminal value, causing heavy encoder computation. To tackle this, we introduce ACV‑Gate, an adaptive candidate evaluation framework that learns to approximate full‑budget counterfactual evaluation and assigns exact evaluations only to the most informative candidates. The core consists of a set‑aware student network trained with terminal advantages and regrets to predict candidate rankings directly; a selective refinement step then evaluates a bounded set containing both Local‑MDL and direct actions; cost‑based thresholds provide explicit control over the average evaluation workload. Experiments on CIFAR‑10 show that ACV‑Gate consistently improves reconstruction quality while drastically reducing candidate evaluations; at 0.20 bpp the primary adaptive configuration raises PSNR by 0.636 dB over LocalMDL with only 2.13 evaluations per image, i.e., 27.60% of the calls required by the Exact‑Full expert. Matched‑candidate comparisons, synchronized GPU measurements, and tests on STL‑10 and a 384×384 transfer further demonstrate stable quality‑computation trade‑offs, especially at low bit rates. In summary, combining terminal‑value learning with selective candidate evaluation offers an effective and controllable mechanism for allocating encoder computation in packet‑constrained generative image communication.
Review