AI Paper Atlas

Research map / Image generation / Autoregressive and latent generation

RADAR: VQ-VAE Decoder of VAR is a Good Student for Restoring Against Degradation by Acceleration

Ziyang Wang, Yue Zhang, Mingdao Wang, Yasen Zhang, Teer Song, Yu Tian, Xueming Li

CVPR 2026 · Image generation › Autoregressive and latent generation · approach: Next-scale and token autoregressive images

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add PAA decoder LoRA adaptation after SCA-Mask (ablation)VAR-d20 · ImageNet-1K 256x256FID: 5.02 -> 2.68Improved
add SCA-Mask attention masking (ablation)VAR-d20 · ImageNet-1K 256x256Latency (ms): 5458±57 -> 2895±88Improved
apply PAA alone to non-degraded VAR (no teacher) (ablation)VAR-d20 · ImageNet-1K 256x256FID: 2.61 -> 2.88Hurt
add PAA does not increase inference cost (ablation)VAR-d20 + SCA-Mask · ImageNet-1K 256x256Latency (ms): 2895±88 -> 2964±84No clear effect
add RADAR (SCA-Mask + PAA) to degraded VARVAR-d24 · ImageNet-1K 256x256FID: 4.29 -> 2.19Improved
add RADAR to degraded VARVAR-d30 · ImageNet-1K 256x256FID: 3.58 -> 2.01Improved
add RADAR vs FastVAR on InfinityInfinity · GenEvalGenEval score: 0.60 -> 0.69Improved
apply RADAR to VARGPT-v1.1 for 1.6x throughputVARGPT-v1.1 · TextVQAaccuracy: 82.0 -> 77.3Hurt
add PAA to existing acceleration methods (FastVAR, ScaleKV, SkipVAR)VAR-d24 · GenEval percentageperformance percentageImproved
PAA decoder-only adaptation instead of retraining the transformer (MVAR)VAR-d24 · GPU hoursGPU hoursImproved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy