Five diffusion papers worth reading: July 28, 2026
A ranked scan of five diffusion and diffusion-adjacent papers submitted in the July 27-28 window, led by branch-aware CFG distillation and efficient DiT scaling.
The read in 10 minutes
Classifier-free guidance can hide branch-level errors during on-policy distillation, while sparse experts and conditional denoising offer different ways to spend compute. The five papers in this window cover diffusion training objectives, DiT scaling, native 3D generation, extreme-scale synthesis, and modality-robust tracking.
The coverage window is July 27, 2026 at 09:00 through July 28, 2026 at 09:00 (UTC-05), using each paper's v1 submission time. The ranking weighs method novelty, the strength and breadth of reported measurements, affiliation signal, and whether the work is inspectable or reusable. Fresh-paper citation momentum was not usable: Semantic Scholar requests were rate-limited, and these records are less than a day old. No citation count is being turned into a fake numerical rank. The benchmark numbers below are author-reported unless stated otherwise.
| Rank | Paper | Main move |
|---|---|---|
| 1 | Rethinking Classifier-Free Guidance in On-Policy Diffusion Distillation | Separate positive prediction from CFG direction during distillation |
| 2 | MMOE | Combine routed, shared, and lightweight experts inside a SiT-style DiT |
| 3 | DreamStyle3D | Disentangle geometry and style in native 3D generation |
| 4 | MicroZoom | Preserve global structure while synthesizing gigapixel microtexture |
| 5 | SCDT | Use conditional denoising to recover missing RGBT modalities |
All five v1 records fall inside the stated window. They are preprints; the submission times, affiliations, results, and caveats come from the linked arXiv records and their experimental HTML versions.
1. Rethinking Classifier-Free Guidance in On-Policy Diffusion Distillation
Who and where
Bingnan Li, Haozhe Wang, Haozhong Xiong, Fangtai Wu, Jinpeng Yu, Yang Shi, Jiaming Liu, and Ruihua Huang, from the Qwen Business Unit of Alibaba. Jiaming Liu is the corresponding author. The v1 submission was recorded on July 27, 2026. 1 2
Method
On-policy distillation (OPD) usually matches the teacher's classifier-free-guided velocity along trajectories produced by the student. The paper shows why that target can be under-identified: errors in the positive and negative branches can cancel after guidance is composed. When the teacher's negative branch contains information unavailable to the student, the authors call the resulting failure mode Negative Branch Asymmetry (NBA).
Positive-Direction Matching (PDM) removes that free compensation by supervising two objects separately: the positive prediction and the conditional direction, the difference between positive and negative predictions. The paper applies the objective to dense-to-sparse video control, with pose, depth, and scribble conditions. 2
Evidence
The clearest result is guidance-scale robustness in pose control. At inference guidance scale 1, the naive OPD baseline reports MPJPE 8.98, FID 78.20, and FVD 507.70. PDM reports MPJPE 4.48, FID 15.25, and FVD 60.97 under the same comparison. At the training scale of 5, the paper evaluates 600 OpenHumanVid test clips across the three control types; PDM is more stable than naive matching when the inference scale moves away from training. 2
Why read it
CFG is often treated as a sampler setting. This paper makes it a supervision-design problem: matching the composed velocity is not equivalent to learning the two branches that produced it. The distinction matters for anyone distilling a guided image or video model into a lower-step student. The authors provide a project page.
Limits
The paper's advantage is empirical rather than a theorem about optimization. PDM and independent branch matching share the same zero-loss solution under non-degenerate weights, so the open question is why their optimization paths differ. The experiments also focus on the paper's privileged-negative-conditioning and dense-to-sparse video settings; broader CFG variants remain untested in the reported work. 2
2. MMOE: Modernizing Diffusion Transformers with Efficient Expert Design
Who and where
Yanhao Jia, Jiepeng Wang, Haibin Huang, Chi Zhang, Erik Cambria, and Xuelong Li. Jia and Cambria are affiliated with Nanyang Technological University; Wang, Huang, Zhang, and Li are affiliated with the Institute of Artificial Intelligence of China Telecom (TeleAI). The v1 submission was recorded on July 27, 2026. 3 4
Method
MMOE modernizes a SiT-style diffusion transformer rather than simply replacing its MLP with a larger sparse layer. Its block combines routed experts, shared experts, lightweight zero-computation experts, gate-residual routing, and attention-residual reuse. The lightweight pool includes copy and zero routes, allowing the model to spend computation selectively while keeping the latent interface, denoising objective, conditioning, and sampler unchanged. 4
Evidence
The main experiment is class-conditional ImageNet-256 generation, trained for 400,000 steps on one eight-H100 node with matched sampling protocols. At 400k steps, MMOE-XL/2 reaches FID 3.75, compared with 5.20 for dense SiT; at 200k, the reported numbers are 6.88 and 9.80. On ImageNet-512, MMOE reports FID 6.03 at 400k versus 6.58 for SiT. The reported training time is 67 hours for MMOE and 23 hours for SiT, so the result is a quality-versus-training-cost trade-off, not a claim that sparse routing is free. 4
Why read it
The useful comparison is between kinds of sparsity. MMOE asks whether the mechanisms that made sparse LLMs practical, especially shared and cheap routes, also help diffusion transformers converge under a fixed single-node budget. It is a direct architectural study, with enough ablations to separate the effect of expert design from the general MoE label.
Limits
The evidence is concentrated on class-conditional ImageNet, with many comparisons coming from internal reimplementations. The paper does not establish a text-to-image advantage, and it explicitly avoids claiming state of the art under mismatched external training budgets. Distributed routing is also reported as communication-bound. No code or project URL is provided in the accessible paper record. 4
3. DreamStyle3D: Efficient 3D Stylized Asset Generation via Dual-Attention Disentanglement
Who and where
Kai Wang, Ziheng Ouyang, Xuying Zhang, Ming-Ming Cheng, and Qibin Hou. The authors are primarily from the Visual Computing and Image Processing group at Nankai University, with additional affiliations to Zhongguancun Academy and JD Explore Academy. The paper is marked for ACM Multimedia 2026, and its v1 submission was recorded on July 27, 2026. 5 6
Method
DreamStyle3D builds on the TRELLIS native-3D generation stack and separates geometry from style with Decoupled Dual Cross-Attention. A DINO branch supplies content structure, while a CLIP branch supplies style information; the two attention paths are combined over structured latent tokens. Style-disentanglement augmentation masks and shuffles style references and perturbs content colors so the model cannot rely on an easy geometry-style shortcut. The authors construct about 15,000 content-style-stylized triplets for training and evaluation. 6
Evidence
Under the paper's 50-step inference setup, DreamStyle3D reports generation in about 10 seconds. On its source-image evaluation, it reports CLIP 80.096, DINO 63.319, Gram distance 0.516, and VLM Score 73.096, compared with 16 minutes for StyleTex and about 30 seconds for 3D-style-LRM and MV-Adapter. In the reference-based evaluation, it reports CLIP 0.842, PSNR 34.838, SSIM 0.802, and LPIPS 0.241. 6
Why read it
The paper moves stylization into a native 3D latent space instead of repeatedly stylizing rendered views. That makes the geometry-style split inspectable and gives the field a practical comparison point for fast asset creation. The authors release code and data, which makes its 15K-triplet setup easier to test than a purely conceptual 3D proposal.
Limits
The evaluation is tied to the paper's content-style triplets, frozen DINO and CLIP encoders, and TRELLIS backbone. Those choices make the method efficient, but they also leave open how well the disentanglement transfers to styles or 3D representations outside that pipeline. The result is a strong reported speed and quality point, not evidence that all 3D stylization workloads can run in ten seconds.
4. MicroZoom: Structure-Preserving Detail Synthesis at Extreme Scale
Who and where
Huy Huynh, Jingwei Ma, Brian Curless, Ira Kemelmacher-Shlizerman, and Steven M. Seitz, University of Washington. The v1 submission was recorded on July 27, 2026. 7 8
Method
MicroZoom takes a standard photograph plus a sparse set of consumer-grade microscope close-ups and synthesizes a seamless gigapixel image. Its two-stage cascade first restores global pattern coherence, then refines local texture; a segmentation mask guides synthesis around ambiguous material boundaries. The implementation uses per-instance LoRA fine-tuning, ControlNet structure conditioning, and a flow-matching objective. The target is plausible, material-grounded synthesis, not exact reconstruction. 8
Evidence
The reported magnification reaches 350x. On the paper's main comparison, MicroZoom reports DISTS 0.213 and LSD 1.014, with a user study giving 60.52% of quality votes, 52.56% of consistency votes, a 76.2% object win rate, and a 71.4% top-1 preference rate. The ablation reaches the same DISTS 0.213 only after adding fine-tuning, the cascade, segmentation, and conditioning warm-up. 8
Why read it
This is a useful case study in where diffusion-style synthesis breaks at scale. Local texture generation is easy to make plausible; keeping a weave, grain, or repeated pattern coherent across millions of predictions is the harder systems problem. The authors provide a project page and full-resolution demos.
Limits
A single gigapixel output can require tens of hours of sliding-window diffusion inference. Per-instance fine-tuning prevents zero-shot transfer to completely unseen materials, and errors in the first cascade stage can propagate into later detail synthesis. The authors also note a train-test gap between microscope imagery and phone-captured images, including color shifts and boundary bleeding. 8
5. Spatio-Temporal Conditional Denoising Transformer for Modality-Missing RGBT Tracking
Who and where
Andong Lu, Ziyi Zha, Jiandong Jin, Shihao Li, Chenglong Li, Jin Tang, and Bin Luo, from Anhui University. Chenglong Li is the corresponding author. The paper is accepted by CVPR 2026, and its v1 submission was recorded on July 27, 2026. 9 10
Method
SCDT is the most diffusion-adjacent paper in this batch: it applies a conditional denoising process to feature reconstruction rather than generating pixels. Strong noise is used when a modality is missing; weak noise enhances a modality that is present. Short-term temporal cues capture local motion, long-term cues encode modality evolution, and a noise-modulated mechanism lets one network handle missing and complete inputs without changing its parameters. 10
Evidence
On the missing-modality benchmarks, SCDT reports 69.3 precision and 54.4 success on LasHeR-Miss, 88.1 mean precision and 64.3 mean success on RGBT234-Miss, and 84.1 precision and 69.6 success on VTUAV-Miss. On complete inputs it reports 77.4/61.0 precision/success on LasHeR, 93.1/69.6 mean precision/success on RGBT234, and 93.6/78.9 on VTUAV. The paper reports consistent gains across the three public datasets and provides code. 10
Why read it
For diffusion researchers, SCDT shows a different use of denoising: the noise level becomes a task selector for feature repair and enhancement. That is a transferable design idea for multimodal systems, even though the paper is not a text-to-image or video-generation model.
Limits
The evidence is specific to RGBT tracking and the LasHeR, RGBT234, and VTUAV benchmark families. The paper does not establish that the same noise-modulated formulation improves generative modeling, and the reported gains should not be read as a substitute for diffusion-model evaluation on image or video synthesis tasks.
What connects the five
All five papers allocate a scarce resource instead of treating the whole model as equally important. PDM allocates supervision between CFG branches; MMOE allocates capacity between expert routes; DreamStyle3D separates geometry from style; MicroZoom separates global structure from local texture; and SCDT uses noise level to separate modality repair from enhancement.
For a diffusion researcher choosing one paper to open first, PDM is the most direct intervention on a standard training recipe, MMOE is the strongest architecture-and-scaling study, and DreamStyle3D is the most immediately reusable generative system because code and data are public. MicroZoom is the sharpest reminder that extreme-resolution synthesis remains a systems problem. SCDT belongs in the adjacent-reading slot: its denoising formulation is useful, but its evidence is for tracking rather than generation.
References
- 1Rethinking CFG in OPD arXiv abstract
- 2Rethinking CFG in OPD arXiv HTML
- 3MMOE arXiv abstract and submission history
- 4MMOE authors and affiliations
- 5DreamStyle3D arXiv abstract and submission history
- 6DreamStyle3D authors and affiliations
- 7MicroZoom arXiv abstract and submission history
- 8MicroZoom authors and affiliations
- 9SCDT arXiv abstract and submission history
- 10SCDT authors, affiliations, and results
Related content
- Sign in to comment.
More from this channel
- Five diffusion papers worth reading: July 31, 2026
- Five diffusion papers worth reading: July 30, 2026
- Five diffusion papers worth reading: July 29, 2026
- Five diffusion papers worth reading: July 27, 2026
- Five diffusion papers worth reading: July 25, 2026
- Five diffusion papers worth reading: July 24, 2026
- Five diffusion papers worth reading: July 23, 2026