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 Digest2026/06/08 23:20:21

Five diffusion papers worth reading this Monday (June 8, 2026)

This extended weekend batch (June 6–8, ~72h) surfaces five preprints across distinct corners of the diffusion pipeline. DAVE (KAIST, ICML 2026) identifies DC-component trajectory lock-in in Flow Matching transformers and cures it training-free, recovering Recall +93% on MS-COCO. TrioPose (CAS) redesigns pose conditioning as a native third stream in MM-DiT with zero-initialized residuals, reaching AP 64.33 (+30% over GRPose) on Human-Art. STREAM (DEEPNOID) introduces Riemannian Flow Matching in UNI feature space for histopathology generation, achieving gFID 6.61 on TCGA-BRCA. AsyncPatch (Google DeepMind) proves the first valid ELBO for joint per-pixel asynchronous diffusion, enabling zero-shot inpainting with FID 15–20% below RePaint. FreeAnimate (Tsinghua SIGS, ICASSP 2026) achieves training-free human animation by using DDIM-inverted preview frames as structural scaffolding, outperforming trained baselines on out-of-domain data.

ArXiv Diffusion Models Digest
ArXiv Diffusion Models Digest2026/06/05 22:57:45

Five diffusion papers worth reading today (June 5, 2026)

Friday's batch selects five papers across distinct sub-areas: UniPixie (SUSTech/UPenn/MIT, CVPR 2026 Highlight) applies Flow Matching to 3D physical property prediction, cutting Young's Modulus error 55.6%; ICM (Warsaw UT/valeo.ai, CVPR 2026) localizes diffusion decision-making to self-attention layers for SOTA bias mitigation; OTP-FM (ICML 2026) reaches scRNA-seq SOTA in 3–5 min training via soft optimal-transport flow matching; ReCache (HSE/Yandex) uses REINFORCE to learn cache schedules, cutting LPIPS 31% on FLUX; SAM-Flow (Tianjin) achieves best CLIP-T/background-LPIPS trade-off via attention-union masking for training-free local edits.

ArXiv Diffusion Models Digest
ArXiv Diffusion Models Digest2026/06/04 22:26:27

Five diffusion papers worth reading today (June 4, 2026)

Thursday's June 4 batch delivers conceptual range: MeshFlow (Meta FAIR, CVPR 2026 Highlight) generates 3D meshes 18× faster than AR via parallel Rectified Flow; OT Flow Matching by Design (Tel Aviv) halves trajectory curvature by treating the Gaussian prior as a design variable; Training-Free SID (U Toronto, CVPR 2026) derives a closed-form patch denoiser that beats 8-hour-trained baselines with zero training time; PILA (CASIA/USTC) injects a physics MoE adapter into frozen Wan and transfers zero-cost to the 14B backbone; and VPE (NTU Singapore) shows a 6.84 dB PSNR gain when AR and diffusion tokens share attention instead of passing through a sequential bottleneck.

これ以上のコンテンツはありません