Five diffusion papers worth reading: July 16, 2026

This issue covers the active Thursday, July 16, 2026 new-listing batch for cs.CV and cs.LG: 174 computer-vision entries and 237 machine-learning entries were listed by arXiv. 1 2

The five papers below are ranked for a diffusion-model researcher by technical novelty, strength of reported evidence, affiliation signal, and immediate usefulness. Same-day citation counts are not yet a meaningful separator for this batch, so they are not used as a fabricated popularity score. The listing date and paper submission date also do not always coincide; each entry keeps the paper-level date shown on its arXiv record.

Paper: Bichuan Zhu, Fulin Li, Jiachao Gong, Jinhua Hao, Kai Zhao, Kun Yuan, Pengcheng Xu, Qiang Wang, Qiao Mo, Yanlong Yuan, Yizhen Shao, Yuxiao Hu, Zixi Tuo, Ming Sun, Chao Zhou, Bin Chen, and Bin Yu, identified in the HTML paper as the LPM Team at Kuaishou Technology. Submitted July 15, 2026. 3

The contribution: LPM treats restoration as a large generative system rather than a short-clip enhancement model. LPM-Image first learns a spatial restoration prior with a DiT, a high-fidelity LPM-VAE, and a Rectified Flow objective. LPM-Video then adds factorized 2D spatial and 1D temporal attention, keeps the 2D VAE, removes fixed temporal positional embeddings, and uses temporal-pyramid inference to connect short windows across arbitrarily long videos. 4

The practical details are unusually important. The Kwai UltraVision training set is described as billion-scale, spans 1,212 content categories, and reports an average Kuaishou Visual Quality score of 4.31 on a 0-5 scale. The LPM-VAE reaches PSNR 27.66 and SSIM 0.807 on the paper's internal reconstruction benchmark. The system also reports a 39.5x end-to-end speedup over its original 25-step pipeline after distillation and deployment optimizations.

Reported evidence: On the paper's internal LPM-Benchmark, LPM-Image reports MUSIQ 73.32, CLIP-IQA 0.7423, and KVQ 4.046. LPM-Video reports MUSIQ 72.11, CLIP-IQA 0.5954, and KVQ 3.63 across the benchmark's video settings. These are author-reported results, not an independent audit. 4

The strongest impact signal is deployment: the paper says LPM processes videos representing approximately 45% of Kuaishou's total viewing time. When combined with Kuaishou's encoding and transmission pipeline, it reports more than 20% lower bitrate at comparable perceptual quality and annual bandwidth savings on the order of hundreds of millions. The words "approximately," "more than," and "on the order of" matter here: the paper does not provide a precise public cost ledger.

Why read it: This is a useful reference for anyone designing a production diffusion restoration stack. The interesting research question is not only how to improve a denoiser, but how data curation, a high-fidelity VAE, flow training, temporal factorization, long-video inference, and serving optimization interact when the output is measured at platform scale.

Resources: No public code link is listed on the arXiv record. The full paper is available from the arXiv HTML page.

Paper: Minh-Quan Le, Armand Comas, Alexandros Lattas, Stylianos Moschoglou, Pedro Velez, Amit Raj, Aaron Germuth, Thabo Beeler, Dimitris Samaras, and Di Qiu. The HTML record identifies a team spanning Stony Brook University and Google DeepMind. Submitted July 14, 2026. 5

The contribution: CO2Jump targets a weakness in masked diffusion multimodal generation: text and image branches may commit to incompatible decisions, and standard masked samplers cannot revise already-unmasked tokens. The proposed Self-Correcting Coupled Markov Jump Process makes one modality's transition rates depend on the other modality's confidence through cross-modal attention. A remasking jump can then retract low-confidence commitments. The sampler is training-free, runs on a frozen masked-diffusion backbone, makes one forward pass per denoising step, and requires no architectural change or auxiliary evaluator. 6

The authors build three joint datasets and benchmarks: JEdit-1M for image editing and understanding, JMaze-200K for maze reasoning, and JNono-200K for nonogram reasoning. On the extended ImgEditBench table, CO2Jump reports overall mAP 0.369, ahead of MDM at 0.354 and MMaDA-Parallel at 0.335. Its target mAP is 0.346, versus 0.330 for Qwen3-VL-8B in the paper's comparison.

The same pattern appears in the synthetic reasoning tasks. CO2Jump reports total joint accuracy of 0.432 on JMaze-Test500 and 0.168 on JNono-Test500. Its out-of-distribution scores are 0.320 for maze and 0.175 for nonogram; the latter is notably above MMaDA-Parallel at 0.113. As the denoising budget rises from 8 to 512 NFE, the paper reports ImgEditBench improving from 1.72 to 1.93 and overall mAP from 0.074 to 0.369.

Why read it: The paper turns remasking from a generic confidence heuristic into a cross-modal correction mechanism. For researchers working on discrete diffusion, multimodal token spaces, or diffusion-based visual reasoning, the key idea is that the sampling trajectory can negotiate commitments instead of treating every early token as irreversible.

Caveat: The image-understanding score uses a pseudo scene graph because every method produces a different target image. The evaluation also uses external models for annotation and scoring. Those choices do not invalidate the result, but they make the benchmark less direct than a fixed-ground-truth image task.

Resources: Project page and full paper. 7

Paper: Eungjune Shim, Hansol Lee, and Eunjung Ju, CLO Virtual Fashion Inc., South Korea. Accepted to ECCV 2026 and submitted July 15, 2026. 8

The contribution: DiffGI addresses a representation problem that ordinary volumetric 3D generators handle poorly: thin shells, open boundaries, and non-manifold surfaces such as garments. Instead of a binary occupancy map, it uses a continuous 2D Truncated Signed Distance Function on a UV-aligned geometry image. Its Differentiable Marching Squares module recovers subpixel boundaries with analytical interpolation, allowing 3D surface losses to backpropagate into the 2D representation. A geometry-aware normal-rendering loss helps preserve sharp surface detail. 9

The resulting DiffGI-VAE compresses the 256x256x4 geometry image into a 32x32x4 latent space. A transformer latent-diffusion model with a flow-matching scheduler then performs label-conditioned, image-conditioned, and occupancy-conditioned 3D generation. The main reconstruction datasets are ABO, with 7,900 commercial furniture scans, and GarmageSet, with 14,801 simulation-ready garments. WARDROBE, with approximately 300 real clothing images, is used for qualitative zero-shot image-to-3D evaluation.

Reported evidence: On GarmageSet, the proposed representation reports Chamfer distance 0.46x10^-3, EMD 0.16, JSD 1.24x10^-3, and normal consistency 0.96. In the single-view image-to-3D comparison, DiffGI reports 23K vertices, CD 1.35x10^-2, F1 0.48, HD 8.42x10^-2, and boundary CD 2.91x10^-2. The paper also reports image-conditioned generation in 1.21 seconds on an RTX 4070 and 8.52 seconds on a MacBook M4 CPU. 9

Why read it: This is a clean example of changing the representation before changing the generator. The paper makes 2D diffusion and flow matching useful for a class of meshes that volumetric pipelines tend to thicken, close, or blur. It is especially relevant to researchers building assets for simulation, apparel, furniture, or UV-aware graphics pipelines.

Caveat: The model does not generate RGB textures or PBR material attributes. The authors also report possible chart seams and rounding around extremely sharp edges, so the compact latent space does not remove the need for mesh-quality checks downstream.

Resources: Project page and full paper. 10

Paper: Andrew Marshall, Xuanang Xu, Xiaoran Zhang, Rui Wang, Lawrence Staib, and James Duncan, with affiliations in Yale University's departments of Biomedical Engineering, Radiology and Biomedical Imaging, and Electrical Engineering. Submitted July 14, 2026. 11

The contribution: The authors argue that a per-pixel denoising objective can smooth the irregular tumor boundaries and heterogeneous morphology that matter for downstream segmentation. Their FD-loss instead aligns the first- and second-order statistics of real and generated images in a frozen feature space. The queue-based implementation keeps 5,000 generated features for a stable distribution estimate while backpropagating through only 16 differentiable 25-step DDIM samples per iteration; another 32 detached samples refresh the queue. Fine-tuning runs for 20,000 steps on one NVIDIA H200. 12

The study covers LiTS and MCT-LTDiag liver CT, plus BraTS brain MRI. The downstream test is more informative than a standalone image-quality score: five-model U-Net ensembles are trained with different augmentation strategies and evaluated on held-out real scans.

Reported evidence: On LiTS, synthetic augmentation raises tumor DSC from 46.9 to 52.3 after FD-loss fine-tuning. On MCT-LTDiag, the corresponding result is 60.9 to 64.5. On BraTS, enhancing-tumor DSC is 46.4 for synthetic augmentation and 47.4 with FD-loss, while necrotic-core DSC moves from 65.0 to 66.9. Image-distribution metrics move in the same direction: LiTS FID falls from 99.20 to 61.37 and FRD from 63.56 to 24.05. 12

Why read it: This paper is a useful reminder that a diffusion model can be judged by the task its samples support, not only by whether its images look plausible or score well on FID. It also gives a practical recipe for making a population-level distribution loss usable with a small medical dataset and a manageable gradient batch.

Caveat: The experiments are 2D, and the benefit depends on the selected feature encoder. MedDINOv3 works best for the abdominal CT settings in the reported ablation, while BiomedCLIP is strongest for the brain MRI setting. The queue reduces, but does not remove, the compute and stability constraints of estimating distribution statistics.

Resources: No public code link is listed on the arXiv record. Start with the paper's methods and tables.

Paper: Tessa Cannon, Michel Tsamados, Petru Manescu, Thomas Newman, Christian Haas, Veit Helm, Weibin Chen, and Randall Scharien. Affiliations span University College London, the Alfred Wegener Institute Helmholtz Centre for Polar and Marine Research, and the University of Victoria. Submitted July 15, 2026. 13

The contribution: RoughNet uses a conditional diffusion U-Net to map six 10-meter Sentinel-2 multispectral observations to locally normalized 1-meter surface-elevation residual fields. The target is not a prettier satellite image; it is a physically meaningful estimate of sea-ice roughness that can support climate modeling and safer travel over landfast ice. The model is trained on airborne LiDAR from two Arctic regions and tested on an unseen third region, Cambridge Bay. 14

The selected configuration combines a cosine noise schedule with a PLMS sampler. On Cambridge Bay, the unseen test region, it reports RMSE 0.089 m, near-zero bias, sigma error 18.36%, NAE 1.17 degrees, and JSD 0.081. The model was trained in approximately 150 minutes on an RTX 3090 Ti; a batch of eight 256x256 patches took about 30 seconds to sample.

Why read it: RoughNet is a good counterexample to the idea that diffusion research is mainly about images for human inspection. Here, the generative prior is used to recover sub-sensor-scale physical structure from coarse observations. The region-held-out setup also makes the generalization question visible instead of hiding it behind random patch splits.

Caveat: The model preserves broad amplitude and orientation better than fine-scale geometry. In Cambridge Bay, the reported zero-normalized cross-correlation falls to 0.1129, and the paper describes missing elongated ridge structures and over-smoothed high-frequency power. The 9-cm RMSE is therefore not evidence that every local ridge is faithfully recovered.

Resources: Code on GitHub, project paper. 15

The most consequential paper is LPM because it documents diffusion restoration as a deployed platform component, not only as a benchmark model. CO2Jump is the most interesting sampler design: it makes cross-modal revision part of the discrete-diffusion trajectory. DiffGI is the strongest representation paper in the group, while FD-loss offers the clearest downstream-task test. RoughNet is the best application paper for readers interested in scientific reconstruction and out-of-domain evaluation.

For a ten-minute reading block, start with LPM's system sections, then read CO2Jump's sampler and ablation table. Read DiffGI if your work touches 3D surfaces or flow-based latent generation; read FD-loss if your evaluation target is downstream utility rather than image similarity; and read RoughNet if you care about physically constrained super-resolution.