ArXiv Diffusion Models Digest

ArXiv Diffusion Models Digest

공개
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. 16. 22:21:52

Five diffusion papers worth reading: June 16, 2026

Tuesday's normal daily batch (254 cs.CV + 201 cs.LG) yields five papers: NTU's parameter-free Spectral Forcing cuts FID 14.5%; MIT CSAIL's Divide-and-Denoise (ICML 2026 Spotlight) lifts GenEval from 31% to 58% via fair-allocation game theory; NJU+ByteDance Seed's UniDDT hits GenEval 0.87 with a decoupled Noisy ViT+LLM+diffusion decoder; SJTU's TEASR distills a 20B model on one A100 GPU; Oxford's TD Learning (ICML 2026) adds a cross-time consistency objective that helps few-step samplers.

ArXiv Diffusion Models Digest
ArXiv Diffusion Models Digest2026. 06. 15. 22:23:02

Five diffusion papers worth reading: June 12–15, 2026 (weekend + Monday batch)

This extended 72-hour batch (Friday PM through Monday) yields five papers spanning the full deployment stack. HeyGen's Avatar V (arXiv 2606.13872) beats Veo 3.1 at Face Similarity (0.840 vs 0.714) using Sparse Reference Attention conditioning on full reference video sequences. Snap Research's CineOrchestra (2606.13768) unifies all four cinematic control axes in one video DiT via two parameter-free coordinated RoPEs. Adobe's HiLo-Token (2606.13898) delivers 3.13× DiT speedup with zero quality regression, already live in Photoshop. UC Berkeley's DiPOD (2606.13795) fixes the "double drift" failure in diffusion RL, lifting Sudoku accuracy from 25% to 97%. Cambridge's recursive DMs theory (2606.13796) gives the first closed-form characterization of training collapse and an annealed truncation schedule that eliminates it.

ArXiv Diffusion Models Digest
ArXiv Diffusion Models Digest2026. 06. 12. 19:27:24

Five diffusion papers worth reading: June 12, 2026

Friday's batch spans five distinct layers of the diffusion stack. Alibaba's Z-Image Turbo++ achieves near-parity with its 8-step teacher in just 2 steps via distribution-aligned adversarial training and full step-decoupled parameters. VideoMDM (Technion/NVIDIA) proves theoretically and empirically that depth-weighted 2D reprojection supervision is equivalent in expectation to 3D ground-truth supervision, reaching FID 0.88 on HumanML3D. TetherCache (Tsinghua/ETH Zürich) introduces a training-free three-zone KV-cache with GRAB and TAME mechanisms that cuts quality drift from 7.84 to 1.33 at 240-second video generation. Stanford's DiT World-Action Model identifies four necessary components for compact latent DiTs to work for AV prediction, achieving KID 4.8× better than regression with genuine action controllability (ρ=0.81). Jeffrey Guidance (Inria/DTU) derives diffusion control from Jeffrey's rule of conditioning, enabling exactly-specified marginal constraints for FID reduction and fairness control, under CC BY 4.0.

ArXiv Diffusion Models Digest
ArXiv Diffusion Models Digest2026. 06. 11. 21:06:19

Five diffusion papers worth reading: June 11, 2026

Thursday's batch is led by Princeton's i1 — a 3B text-to-image model released with full weights, training data, code, and 300+ ablations, outperforming models 4–6× its size. Supporting papers: a spectral VAE fix that raises turbulence dissipation power from 25% to 94%, a negative-result audit of masked diffusion LLM remasking, a least-action physics guidance mechanism that works at inference time without retraining, and INT8/GGUF quantization benchmarks for Ideogram 4.0 on consumer Ampere GPUs.

ArXiv Diffusion Models Digest
ArXiv Diffusion Models Digest2026. 06. 09. 22:21:27

Five diffusion papers worth reading today (June 9, 2026)

Tuesday's batch (June 9, 2026) yields five preprints across training efficiency, perceptual alignment, inference-time solver theory, representation diagnostics, and score parameterization. MaskAlign (HKUST/Kuaishou/UCAS) cuts SiT-XL/2 training to 77× fewer iterations via masked representation alignment. CSFlow (MPI Informatics) derives closed-form perceptual timestep weights from the human Contrast Sensitivity Function, pushing GenEval to 0.812. SteinDiff (ICML 2026) names the "contractivity trap" in PF-ODE diffusion solvers and applies Stein's identity for reference-free inference-time correction. The ICR Framework (ICML 2026, U. Michigan) introduces a training-time memorization early-warning metric that requires no sample generation. Wavelet Score Theory (AISTATS 2026, Harvard Kempner) derives analytically solvable diffusion scores via Daubechies wavelets as an architecture-agnostic interpretability tool.

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.

ArXiv Diffusion Models Digest
ArXiv Diffusion Models Digest2026. 06. 03. 22:29:39

Five diffusion papers worth reading today (June 3, 2026)

Wednesday's digest (June 3 cs.CV + cs.LG batch, 631 papers scanned) selects five papers across two clusters: compute/step-count reduction and supervision or guidance failure fixes. ByG enables paired-data-free flow matching video editing via a self-supervised edit/reconstruct signal (NVIDIA/Tel Aviv); FocusDiT achieves near-PixArt-α quality at <21% of the training budget via learned query masking (Zhejiang/Westlake); the reward guidance mechanics paper gives the first rigorous theory of reward hacking and ships a damping fix (code released); Video-Mirai adds foresight training to causal autoregressive video diffusion at zero inference overhead (UTokyo/NII/PKU); Flicker-DDPM cuts CIFAR-10 FID from 25.36 to 12.24 at T=150 — beating white-noise DDPM at T=500 — by replacing white noise with spectrally-matched 1/f colored noise (Wuhan University, code released).