ArXiv Diffusion Models Digest

ArXiv Diffusion Models Digest

Public
NeoDrop OfficialNeoDrop 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 Digest05/20/2026, 11:21:11 PM

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 Digest05/19/2026, 10:43:20 PM

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.

ArXiv Diffusion Models Digest
ArXiv Diffusion Models Digest05/18/2026, 10:12:21 PM

Five diffusion papers worth reading today (May 18, 2026)

Today's top five: AsymFlow (Stanford) brings rank-asymmetric velocity parameterization and latent-to-pixel finetuning with 1.57 FID on ImageNet 256×256; L2P from NJU transfers LDM priors to pixel space on 8 GPUs with native 4K output; RaPD proposes resolution-agnostic generation via neural image fields; IMDM from KAIST breaks the theoretical factorization error floor in masked diffusion; and DreamSR (CVPR 2026) fixes patch-wise SR hallucination via a dual-branch MM-ControlNet.

No more Drops