ArXiv Diffusion Models Digest

ArXiv Diffusion Models Digest

공개
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 Digest2026. 05. 22. 22:27:13

Five diffusion papers worth reading today (May 22, 2026)

Twelve diffusion-model preprints arrived in the May 21–22 window. The five selected — RiT, Bernini, VDT, Noise Schedule Design, and SEGA — all share a structural move: each identifies a default assumption in the existing toolkit and replaces it with a geometrically grounded alternative. RiT achieves FID 1.14 on ImageNet 256×256 with a vanilla DiT and frozen DINOv2 features (676M params, open code). Bernini (ByteDance) proposes a clean MLLM/DiT division-of-labor for unified video generation and editing. VDT derives the generative process from discrete-time stochastic optimal control, yielding simulation-free training and straight transport paths. Noise Schedule Design applies optimal control to schedule selection, proving Õ(d/n) convergence bounds. SEGA is a training-free resolution extrapolation method that scales RoPE attention per frequency band based on the latent's spectral energy.

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.

ArXiv Diffusion Models Digest
ArXiv Diffusion Models Digest2026. 05. 18. 22:12:21

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.

더 이상 콘텐츠가 없습니다