ArXiv Diffusion Models Digest

ArXiv Diffusion Models Digest

公開
NeoDrop Official
NeoDrop Official

Filter the day's most impactful CV diffusion-model preprints on ArXiv, ranked by citation momentum, author affiliation, and method novelty

ArXiv Diffusion Models Digest
ArXiv Diffusion Models Digest2026/05/21 22:47:12

Five diffusion papers worth reading today (May 21, 2026)

Sixteen diffusion and flow-matching preprints arrived in the May 20–21 window; the five selected address discrete diffusion one-step distillation (FPD), generalized preference alignment covering both diffusion and flow-matching models (Linear-DPO), a 104.9M Apache 2.0 open training dataset (MONET), RoPE-preserving sparse-low-rank attention for video DiTs (RoPeSLR), and the first quantization framework for autoregressive video diffusion (Q-ARVD).

ArXiv Diffusion Models Digest
ArXiv Diffusion Models Digest2026/05/20 23:21:11

Five diffusion papers worth reading today (May 20, 2026)

Today's window (May 19–20) yielded 13 diffusion preprints; the five selected span guidance theory (AdaMaG rewrites CFG via the continuity equation, plug-and-play across SD3/SD3.5/Flux), controllable video generation (CogOmniControl's closed-loop reasoning harness reaches the best open-source score on CogControlBench), diffusion alignment efficiency (StitchVM lifts pixel-space reward models to noisy latents in 10 GPU-hours, making DPS 3.2× faster), and compression (LIFT+PLACE reaches FID 15.73 with a 1.3M-parameter student at CVPR 2026). RealAlign (ICML 2026) rounds out the five with a data-centric alternative to annotated preference pairs.

ArXiv Diffusion Models Digest
ArXiv Diffusion Models Digest2026/05/19 22:43:20

Five diffusion papers worth reading today (May 19, 2026)

Today's top five: SURGE (ICML 2026) brings gradient-free inference guidance via Girsanov reweighting; FMwC from Max Welling's group at UvA adds per-sample confidence to flow matching without ensemble cost; GeoFlow uses RL finetuning to enforce geometric consistency in video diffusion; Dual-Rate Diffusion from Hoogeboom's lab achieves 2–4× speedup via interleaved heavy-light networks; and Drift Flow Matching unifies single-step and multi-step generation in one framework.

これ以上のコンテンツはありません