Five diffusion papers worth reading: July 25, 2026

The July 24 ArXiv new-listing pages surface five papers that attack different failure modes in diffusion-based vision systems: regional leakage, motion blur, identity mixing, one-step edit control, and multi-view hallucination. The ranking below favors method novelty, strength and breadth of the reported evidence, author-affiliation signal, and public resources. Same-day citation momentum is not usable here: the available citation records were either zero or unavailable for papers only a day or two old.

One timestamp caveat matters. All five papers were surfaced by the active July 24 cs.CV listing, but their v1 dates are not identical: MagicMakeup, WhereEdit, and GroupVideo are dated July 23, while RealVDeblur and 3D-GIMP are dated July 22. The latter two are included as high-signal listing entries, not relabeled as strict July 24 submissions.

Authors and affiliation. Ziyi Wang, Siming Zheng, Yang Yang, Shusong Xu, Hao Zhang, Bo Li, Changqing Zou, and Peng-Tao Jiang. The paper lists Zhejiang University's State Key Lab of CAD&CG and vivo's BlueImage Lab. v1: July 23, 2026. 1

What changes. MagicMakeup treats makeup transfer as a spatial-control problem rather than a full-face style-transfer problem. Token-Aligned Region Gating maps pixel masks onto the DiT token grid and applies region-specific logit gating, aiming to stop makeup attributes from leaking into non-target areas. Cross-Modal Perception Guidance separately aligns text and image features for the attributes to transfer and the identity features to preserve. The data pipeline creates identity-consistent, region-labeled 1024 x 1024 pairs and packages them as the MakeupHQ Bench.

Evidence. The authors start from 50,000 real makeup faces, retain about 3,000 standardized images, and report 6,772 final region-labeled pairs: 2,392 face, 2,224 eye, and 2,156 lip pairs. On MakeupHQ-Synthetic, the reported DINO-I is 0.660, CLIP-I is 0.816, Face-ID is 0.730, and FID is 59.238. On MakeupHQ-Real, the corresponding reported values are 0.623, 0.799, 0.706, and 76.273. In an ablation, Face-ID rises from 0.870 to 0.922 for eyes, from 0.879 to 0.971 for lips, and from 0.805 to 0.858 for full-face transfer when the components are added. These are author-reported comparisons, not independent evaluations. 1

Why read it. The useful idea is not the makeup application itself. It is the decision to enforce spatial locality at the attention-logit level while separately representing transfer and preservation. That makes the paper relevant to any DiT editor where a strong semantic change must coexist with a protected background or identity region.

Limitations. The evidence is concentrated on a newly constructed benchmark and the paper's own synthetic and real test splits. The reported FID and identity metrics are not directly comparable across every editing setup. The paper does not provide a separate GitHub repository in the accessible record.

Resources. Paper and HTML and the MagicMakeup project page.

Authors and affiliation. Renbiao Jin, Mingxin Yang, Yutian Chen, Junhao Zhuang, Xin Cai, Mulin Yu, Linning Xu, Wenxian Yu, Danping Zou, Shi Guo, and Tianfan Xue. The paper lists Shanghai Jiao Tong University, Shanghai AI Laboratory, CUHK MMLab, CPII under InnoHK, and Tsinghua University. The accessible record does not provide a reliable author-by-author mapping for those numbered affiliations. v1: July 22, 2026. 2

What changes. RealVDeblur reuses a pretrained DiT-based video generation model as a spatiotemporal prior for real-world deblurring. Its frame-wise latent encoding keeps frame-dependent blur from being collapsed by temporal VAE compression. Single-step distribution-matching distillation reduces multi-step sampling to one generator, while a training-free Temporal Window Mask is intended to prevent positional-encoding extrapolation from destabilizing long videos. The training data, OmniBlur, combines 3D Gaussian Splatting scenes with high-frame-rate video to model camera motion, object motion, defocus, and compound blur.

Evidence. The paper reports results on BSD, RealBlur, RSBlur, FEVD, and RWBI, with PSNR, SSIM, LPIPS, FID, MUSIQ, NIQE, and temporal optical-flow error. On BSD, the full model reports PSNR 28.76, SSIM 0.884, LPIPS 0.118, FID 9.1, and tOF 1.974. On RealBlur it reports PSNR 27.55, SSIM 0.852, LPIPS 0.100, and tOF 1.827. The reported full-model runtime is 0.15 seconds per frame. The authors also test downstream 3D reconstruction, reporting PSNR 25.41 and LPIPS 0.1628 for real camera motion on DeblurNeRF. These figures are author-reported and include the paper's own data synthesis and evaluation choices. 2

Why read it. This is the strongest deployment-oriented paper in the batch. It connects three questions that are often evaluated separately: whether a diffusion prior restores perceptual detail, whether one-step distillation preserves temporal behavior, and whether the result helps a downstream 3D task. The 3DGS-derived blur synthesis is also a useful recipe for building less artificial restoration data.

Limitations. The main comparisons are still internal benchmark evaluations, and the authors' own evidence shows the usual perceptual-fidelity tradeoff rather than eliminating it. The paper's v1 date is July 22, so it is a listing-date inclusion rather than a strict July 24 v1 submission.

Resources. Paper and HTML and the RealVDeblur project page.

Authors and affiliation. Ming Hu, Mingyu Dou, Jianfu Yin, Miaomiao Zhang, Cong Hu, Yao Wang, Bingliang Hu, and Quan Wang. The accessible HTML record shows four numbered affiliations but does not expand their institution names, so no institution is inferred here. v1: July 23, 2026. 3

What changes. WhereEdit reframes one-step image editing as localized adaptive editing. Its Amplified Conditional Transport increases target attraction instead of relying on a simple source-target prediction difference. Its Attention-Guided Automatic Local Edit Mask, or AutoMask, uses prompt differences and UNet cross-attention to discover a soft edit region without an external segmentation model, extra labels, or additional training. The paper is best read as diffusion-adjacent one-step editing: the central claim is about how a one-step T2I editor uses its internal attention and latent transport.

Evidence. On PIE-Bench, which the paper describes as 700 instruction-based examples across 10 categories at 512 x 512, the reported one-step WhereEdit result is PSNR 23.96, MSE 8.05, CLIP-Whole 25.21, CLIP-Edited 22.66, 0.59 seconds runtime, and 7,436 MiB VRAM. The reported ground-truth-mask variant reaches PSNR 30.21 and MSE 1.61. With the default amplification coefficient k = 2.0, the paper reports PSNR 23.96 and CLIP-Edited 22.67; increasing k from 0.5 to 4.0 raises CLIP-Edited from 21.92 to 23.10 while lowering PSNR from 25.65 to 21.59. 3

Why read it. WhereEdit isolates a failure mode that is easy to hide behind aggregate editing scores: a model can make the requested semantic change while editing the wrong pixels. The combination of automatic localization and adjustable target attraction gives researchers a concrete way to probe that tradeoff in one-step editors.

Limitations. The affiliation metadata is incomplete in the accessible record. The strongest localization result uses a ground-truth mask, while the practical setting must infer AutoMask. All runtime, memory, and quality claims are author-reported on PIE-Bench.

Resources. Paper and HTML and the WhereEdit code repository.

Authors and affiliation. Xinyang Song, Libin Wang, Jianxin Sun, Qi Li, Dandan Zheng, JingDong Chen, and Zhenan Sun. The paper associates the University of Chinese Academy of Sciences, the New Laboratory of Pattern Recognition and State Key Laboratory of Multimodal Artificial Intelligence Systems at the Institute of Automation, Chinese Academy of Sciences, with Ant Group. Qi Li is the corresponding author. v1: July 23, 2026. 4

What changes. GroupVideo targets a weakness in identity-customized video generation: concatenating several face references often creates identity blending and unnatural copy-paste motion. Built on a Video Diffusion Transformer, it uses visual alignment to jointly encode multiple identity references and a semantic perceiver to improve motion semantics. An ID localization module adds spatial guidance, bounding-box constraints, and mask regularization so the model can focus its capacity on facial regions. The authors also construct a 20,000-video multi-identity dataset.

Evidence. In the reported two-identity setting, GroupVideo reaches FID 147.228, Aesthetic Quality 0.611, Dynamic Degree 0.821, FaceSim 0.636, and CLIPScore 32.035. The listed ConsisID baseline reports FID 162.547, FaceSim 0.634, and CLIPScore 31.245. In the paper's ablation, removing visual alignment reduces FaceSim from 0.733 to 0.498, while removing the semantic perceiver reduces CLIPScore from 31.95 to 28.55. These results are from the authors' benchmark and ablations; no independent replication is provided in the accessible record. 4

Why read it. This is a useful test of whether reference conditioning scales beyond the single-person assumption built into many identity adapters. The paper's value is the decomposition of the problem into identity alignment, motion semantics, and spatial localization, which makes failure analysis more concrete than a single face-similarity score.

Limitations. The paper claims support for an arbitrary number of identities, but the quantitative comparison shown here is centered on the two-identity setting. No public code, project page, or demo link was present in the accessible record. The v1 date is July 23, while the paper was selected from the July 24 active listing.

Resources. Paper and HTML. No public implementation link was identified in the paper record.

Authors and affiliation. Xuening Tian, Dieter Schmalstieg, and Shohei Mori, University of Stuttgart. v1: July 22, 2026. 5

What changes. 3D-GIMP performs one generative inpainting pass on a key reference view of a 3D Gaussian Splatting scene, then propagates the completed texture through the other views with a 3D-aware PatchMatch procedure. A Poisson formulation completes depth and helps align the propagated patches. The design deliberately moves stochastic generation to one view and uses scene geometry for the rest, targeting the multi-view inconsistency and hallucination drift that can appear when every view is diffused independently.

Evidence. On IMFine, the reported 2DGS variant reaches PSNR 19.09, LPIPS 0.248, FID 53.11, and six minutes of inpainting time. On 360-USID, its reported 2DGS result is PSNR 15.02, LPIPS 0.387, FID 57.38, and 6.8 minutes. In the comparison with GEN3C on IMFine, the authors report PSNR 19.81, SSIM 0.977, and LPIPS 0.238 for their 2DGS setup. The ablation improves from PSNR 17.48 and FID 122.47 without warping or a mapping field to PSNR 19.09 and FID 53.11 with both. PSNR and LPIPS are computed on the object mask, and the time measurement excludes 3DGS re-optimization. 5

Why read it. 3D-GIMP is the batch's clearest example of using diffusion as a prior rather than as the whole pipeline. If the target application has reliable geometry and many views, the paper asks a productive systems question: which parts need generative sampling, and which parts should be solved by correspondence and reconstruction instead?

Limitations. The reference-view choice and the scene-specific depth scale affect results. The authors note possible artifacts under extreme occlusion and complex reflections. Because the v1 date is July 22, this is a high-signal active-listing inclusion rather than a clean strict-window hit. No public code or project page was identified in the accessible record.

Resources. Paper and HTML.

Read MagicMakeup first if your work involves localized DiT editing and identity preservation. Read RealVDeblur for a deployment-oriented look at one-step video restoration, long-window stability, and downstream 3D reconstruction. WhereEdit is the most immediately reproducible entry because its code is public and its one-step control tradeoff is explicit. GroupVideo is the relevant paper for multi-person identity conditioning, while 3D-GIMP is the systems counterpoint: generate once where the prior is valuable, then use geometry to preserve consistency.

The common thread is allocation. These papers move computation and control to the place where diffusion is most useful: region tokens, frame-wise latents, identity references, or a single geometrically anchored view.