Artifact-Compensated Training for JPEG-to-JPEG-AI Recompression Avoids 29x Slowdown
ssh4net · x · 2026-11-12
A new paper tackles recompressing JPEG images into the JPEG AI format. The authors show that directly applying JPEG AI (trained on uncompressed images) to JPEG inputs preserves the original JPEG artifacts in the output. Using an external artifact-removal neural pre-filter fixes this but increases processing time by 29x. The paper instead proposes artifact-compensated training for JPEG AI, combining pre-filtering against target distortions with efficient compression in a single step.
More from Research
- "Hard to tell a 92% model from a 99% model": the bottleneck is evals — chris_j_paxton · 2026-11-12
- Smartwatch PPG + accelerometer signals reframed to detect hypertension signs — CurieuxExplorer · 2026-11-12
- Open-source personalized mRNA cancer vaccine for a pancreatic cancer patient — sumanthd17 · 2026-11-12
- SteerDistill: teacher-guided quantization keeps W4A16 agents within 1 point of BF16 — _akhaliq · 2026-11-12
- MIMIC and BMW train video-action model hitting up to 100% success on industrial tasks — pess_r · 2026-11-12
- Paris seminar: Monte Carlo methods bring exactness guarantees to diffusion model sampling — kastnerkyle · 2026-11-12