Five diffusion papers worth reading: August 27, 2026 — streaming 4D control, texture grafting, and generative backdoor defense

The Thursday, August 27 arXiv New submissions pages supply the candidate surface for this issue: 90 new entries in cs.CV and 105 in cs.LG. 1 2

Each selected paper carries a v1 stamp of August 26, 2026. The local channel window runs from August 26 at 09:00 through August 27 at 09:00 (UTC-05:00), which corresponds to August 26 14:00–August 27 14:00 UTC. The five abstracts below fall inside that stamp range.

Same-day citation counts and complete social-momentum scores were not available for these fresh submissions. Hugging Face Daily Papers for August 26–27 lists other work with upvotes, but does not give a usable ranking signal for this five-paper set, so the order below uses method novelty, author-reported evidence, affiliation signal, and public inspectability. 3 4

The five-paper scan

Rank	Paper	Main mechanism	Strongest reported evidence	Read first if you work on...
1	4DStreamCtrl	Unified 3D point-track control plus 4-step causal streaming	Teacher EPE 5.29 / LPIPS 0.404 on DAVIS; student 20.6 FPS at 480p	Interactive video, 4D control, streaming diffusion 5 6
2	GraftSR	Dual-mask identical-instance texture grafting for one-step SR	LPIPS 0.1583 on TexRefSR-Eval Syn (−20.2% vs. next best)	Reference-guided restoration and generative SR 7 8
3	DEFUSE	Conditional diffusion reconstruction as SSL backdoor score	AUPRC 0.81–0.96 across seven ImageNet attack settings	Encoder security and generative priors 9 10
4	InteractGesture	Progressive chunk guidance over frozen gesture diffusion	Progressive FGD 0.431, 6.335 cm avg. control error on BEAT2	Streaming motion control and avatar diffusion 11 12
5	SBP fair-face guidance	One-shot early latent nudge from late-stage demographic boundaries	Gender FD 0.051→0.001 (~98% reduction) on CelebA-HQ LDM	Fair generation and inference-time latent editing 13 14

1. 4DStreamCtrl: interactive 4D video in a stream

Authors and institution. Shiqian Li (Peking University; Tencent Hunyuan), Chenguo Lin (Peking University), Zhiguang Liu, Yu Tang, Jiarong Ou, and Rui Chen (Tencent Hunyuan), and Yixin Zhu (Peking University; corresponding). v1: August 26, 2026. Categories: cs.CV, cs.AI. 5 6

What it does. 4DStreamCtrl turns camera motion, object trajectories, and depth into one 3D point-track interface. A lightweight Geometric Motion Head rasterizes those tracks onto the VAE latent grid and feeds a Wan2.2 TI2V-5B video diffusion backbone. The same pathway supports joint camera–object control, depth editing, and motion transfer. Because the motion encoder is temporally separable, the authors distill a 50-step bidirectional teacher into a 4-step causal student that keeps memory independent of video length. 6

Technical insight. Image-plane trajectory controllers miss depth and occlusion; offline 3D controllers usually fix clip length. 4DStreamCtrl keeps the geometry explicit and still streams: OpenVidHD-Motion3D supplies about 0.4M in-the-wild clips with 32×32 3D tracks and per-frame cameras, LoRA adapts the DiT in two resolution stages, and self-forcing / DMD-style distillation yields the causal student. 6

4DStreamCtrl pipeline from 3D tracks through a geometric motion head into Wan 2.2 video diffusion — *The authors' figure routes extracted or user-specified 3D tracks into 2D and depth motion embeddings, then concatenates them with noisy latents inside a Wan2.2 TI2V-5B DiT. 6*

Evidence. On 30 DAVIS validation videos with joint object and camera control, the teacher reports EPE 5.29, LPIPS 0.404, SSIM 0.479, and PSNR 16.04 at 0.84 FPS. Against MotionStream on the same Wan2.2-5B backbone, replacing 2D tracks with 3D tracks cuts EPE by about 33% (7.86 → 5.29) and LPIPS by about 5% (0.427 → 0.404). The causal student reaches 20.6 FPS at 480p with EPE 5.48, within 4% of the teacher, and ahead of MotionStream Causal on Wan2.1-1.3B (16.7 FPS) and Wan2.2-5B (10.4 FPS). The paper also reports coherent 350-frame rollouts (14.6 s at 24 FPS). PSNR is the one metric the authors do not lead (16.04 vs. 16.61 for a competitor), which they attribute to pixel-exact reproduction penalizing plausible variation. 6

Code/demo. The paper lists a project page. The retrieved abstract and HTML do not establish a public code release. 5 15

Caveat. Track quality depends on monocular 3D annotation (SpatialTrackerV2 on OpenVid-1M). Extreme camera baselines can still show depth-scale inconsistency, and the streaming student uses a fixed local attention window. Numbers are author-reported on the paper's DAVIS protocol and single-GPU throughput setting. 6

Why it matters. Controllable video is splitting into camera-only, 2D-drag, and offline-3D stacks. 4DStreamCtrl collapses those into one track representation and still hits interactive latency. Read it if your bottleneck is online 4D control, causal distillation, or building motion datasets that carry depth instead of only image-plane paths.

2. GraftSR: grafting the same object's texture

Authors and institution. Qifan Yu, Haoran Bai, Zongyao He, Weijie He, Sibin Deng, and Ying Chen (Taobao & Tmall Group of Alibaba; Ying Chen corresponding), and Honggang Qi (University of Chinese Academy of Sciences). v1: August 26, 2026. Category: cs.CV. 7 8

What it does. GraftSR is a one-step generative super-resolution model that takes a low-quality image plus a reference photo of the same instance. Dual-mask guidance separates two jobs: mask-modulated reference tokens decide which authentic textures to pull from the reference, and region-aware semantic tokens decide where those textures land on the target. The stack is an MMDiT trained with dual-noise LQ tokens, MSE + DISTS + adversarial loss. The authors also release TexRefSR-141K (about 141K reference tuples / 61K instances) and TexRefSR-Eval (300 cases: 250 synthetic, 50 real). 8

Technical insight. Ordinary diffusion SR hallucinates plausible texture when the low-quality input is underdetermined. Reference-based SR usually assumes brittle spatial alignment. GraftSR refuses that assumption: complementary masks from a VLM (Qwen2.5-VL) and SAM3 mark what to copy and where to place it, so the model can transfer fabric, print, and surface detail without a hard warp. 8

GraftSR dual-mask reference guidance over an MMDiT one-step SR backbone — *Region-aware semantic tokens and mask-modulated reference tokens enter the MMDiT alongside dual-noise LQ tokens for one-step restoration. 8*

Evidence. On TexRefSR-Eval Syn, GraftSR reports LPIPS 0.1583, DISTS 0.0986, PSNR 30.1060, SSIM 0.8494, MUSIQ 69.8535, NIQE 4.9346, CLIPIQA 0.7275, and MANIQA 0.5270. The LPIPS figure is a 20.2% reduction versus the next-best baseline (ODTSR). On the Real split, the paper reports NIQE 3.9881, MUSIQ 73.0045, CLIPIQA 0.6409, MANIQA 0.6507, and VLM scores PQ 3.980 / TC 2.840 / Overall 3.410. Against Gemini-3-Pro and GPT-Image-2 on Syn PSNR, the table lists 30.11 vs. 26.79 vs. 18.33. 8

Code/demo. The paper lists a GraftSR project page. The retrieved abstract and HTML do not establish a public code or dataset download URL beyond that page. 7 16

Caveat. Gains are measured on a new identical-instance protocol the authors introduce. Transfer to classical multi-view reference SR benchmarks is still an open question, and mask quality depends on the VLM/SAM pipeline used at data construction and inference. 8

Why it matters. Generative SR keeps winning perceptual tables while losing the object’s real surface. GraftSR makes that surface an explicit conditioning problem rather than a hope that the prior will guess right. Read it if you care about product imagery, identity-preserving restoration, or building reference datasets that pair the same physical instance across views.

3. DEFUSE: diffusion priors that expose SSL backdoors

Authors and institution. Tuo Chen (Southeast University; Ant Group), Jie Gui (Southeast University; Purple Mountain Laboratories; corresponding), Minjing Dong (City University of Hong Kong), Lanting Fang (Beijing Institute of Technology), Ju Jia (Southeast University), Benlei Cui (Alibaba Group), and Jian Liu (Ant Group). v1: August 26, 2026. Category: cs.CV. Comment: accepted at ACM Multimedia 2026. 9 10

What it does. DEFUSE detects backdoored inputs for self-supervised visual and vision-language encoders. A frozen suspect encoder yields a representation; a lightly fine-tuned conditional diffusion model reconstructs an image from that representation; a frozen reference encoder (DINOv2) scores cosine similarity between the original and the reconstruction. Low semantic consistency marks a backdoor. The method does not need the defender to know the attack recipe or hold matched clean in-distribution labels for every victim. 10

Technical insight. Pixel-faithful reconstruction from a pooled SSL feature is intractable, so DEFUSE relaxes the objective to semantic reconstruction and leans on a pretrained diffusion prior (SDXL works best in the ablations). Clean features reconstruct into content that stays close in DINOv2 space; backdoored features often map toward the attacker’s target class or into meaningless images. 10

DEFUSE reconstruction pipeline from a frozen suspect encoder through conditional diffusion to a reference-encoder backdoor score — *Trainable projection and selected cross-attention modules condition a mostly frozen diffusion denoiser; the reference encoder turns reconstruction consistency into a detection score. 10*

Evidence. On ImageNet attack suites spanning data poisoning and training manipulation, DEFUSE reports AUPRC 0.81 (SSLBKD), 0.84 (CTRL), 0.88 (BLTO), 0.96 (CLIP Backdoor), 0.89 (BadEncoder), 0.81 (DRUPE), and 0.90 (BadCLIP), with the highest AUPRC in each column of the main comparison table. As a purifier that resamples from t = 0.3T, it cuts CLIP-Backdoor ASR from 95.2% to 8.0% and BadCLIP ASR from 88.9% to 0.0%. Under a white-box adaptive attack that maximizes reconstruction similarity, AUROC falls to 0.78 and rises back to 0.90 after adding Gaussian noise with variance 0.1. Fine-tuning the conditional diffusion path improves AUROC by 0.09–0.17 over training-free reconstruction baselines, and 20 inference steps already saturate detection quality. 10

Code/demo. Source code is listed at https://github.com/jsrdcht/DEFUSE. 9 17

Caveat. Detection still needs a public generative model, a reference encoder, and compute for reconstruction. Threshold selection uses Youden’s J and a fitted linear map from DINOv2 similarity (R² = 0.67). The 1%-poison / 10%-filter setting remains hard for most defenses; DEFUSE’s strongest reported filter result in that regime is on CLIP Backdoor. 10

Why it matters. Most diffusion papers ask the prior to synthesize. DEFUSE asks the prior to audit a representation. That is a concrete security use of generative models for researchers who ship or consume SSL encoders, and the ACM MM acceptance plus public code make the paper unusually easy to inspect on day one.

4. InteractGesture: streaming spatial control for co-speech motion

Authors and institution. Ekkasit Pinyoanuntapong (Meta; University of North Carolina at Charlotte), Ajinkya Deogade, Paul Streli, Wenjing Zhang, Joanna Materzynska, Vittorio Ferrari, and Jie Shen (Meta), and Pu Wang (University of North Carolina at Charlotte). v1: August 26, 2026. Category: cs.CV. Comment: ECCV 2026 Workshop — Interactive Social Avatars. 11 12

What it does. InteractGesture is an inference-time controller for a frozen co-speech gesture diffusion model (GestureLSM on BEAT2). At selected reverse steps, the method decodes the current target latent through a differentiable RVQ-VAE, measures a masked spatial loss against user joint targets, and backpropagates into the latent rather than into model weights. Progressive Chunk Guidance keeps a staggered window of editable chunk latents so future constraints can still correct earlier trajectories during streaming generation. 12

Technical insight. Sequential chunk inference freezes finished chunks, so a late pointing target cannot fix an early wrist path. Synchronous guidance can fix that offline but needs the full audio sequence. Progressive Chunk Guidance is the streaming middle path: new chunks enter with offset Δi = 1, context frames refresh from the latest guided estimate, and gradients cascade across chunk boundaries inside a fixed latency budget. 12

InteractGesture architecture with progressive chunk forcing guidance over frozen gesture latents — *Spatial guidance losses on body-part tokens update active chunk latents while previous-chunk context and staggered denoising steps keep the stream causal. 12*

Evidence. On BEAT2, across 15 control settings (1/2/5 keyframes × four joints and all-four), Progressive Chunk Guidance reports FGD 0.431, beat consistency 0.762, diversity 11.760, trajectory error 0.267, location error 0.161, and average control error 6.335 cm. Synchronous Chunk Guidance is the offline accuracy upper bound on control error (4.669 cm) but trails Progressive on FGD (0.442). Sequential guidance lands at 11.701 cm average error. Inverse kinematics hits 19.973 cm with FGD 1.043; a chunk-wise ControlNet baseline reaches only 46.159 cm average error. Ground-truth beat consistency and diversity are 0.703 and 11.970. 12

Code/demo. The paper lists a project page. The retrieved abstract and HTML do not establish a public code release. 11 18

Caveat. Guidance concentrates in late sampling steps because early decoded joints are noisy. Dense multi-joint control needs gradient step-size normalization, or naturalness collapses. Results are tied to GestureLSM, BEAT2, and the authors’ K_late = 30 / K_post = 40 schedule. 12

Why it matters. Streaming diffusion control is not only a video problem. InteractGesture shows the same chunk-boundary failure mode in motion latents and fixes it without retraining the generator. Read it if you build interactive avatars, test-time guidance, or any chunked diffusion sampler that must accept mid-stream spatial edits.

5. SBP: learn the boundary late, nudge once early

Authors and institution. Subir Kumar Parida, R. S. Sengar, and Swati Hiremath (Bhabha Atomic Research Centre, Mumbai), and Rajbabu Velmurugan and Ketan Kotwal (Department of Electrical Engineering, IIT Bombay). v1: August 26, 2026. Category: cs.CV. 13 14

What it does. Semantic Boundary Predictor (SBP) is a training-free-for-the-backbone fairness method for latent diffusion face generators. The authors observe that late denoising latents separate demographic classes more cleanly, while early noisy latents still move easily. SBP learns linear demographic boundaries from late-stage latents of the biased model’s own samples, then applies a single offset z_T ← z_T ± δ·b before reverse diffusion. No LDM fine-tuning and no external balanced training set are required. 14

Technical insight. Iterative guidance methods pay a cost at every reverse step. SBP decouples analysis from intervention: the boundary is estimated where class structure is sharp, then applied once where the sampler is still flexible. PCA projection keeps the boundary learning cheap in high-dimensional latent space. 14

SBP transfers a late-stage demographic boundary into a one-shot early latent nudge before reverse diffusion — *A boundary vector b learned at the clean end of the reverse process is added once to the initial noisy latent, after which the frozen LDM samples as usual. 14*

Evidence. On a CelebA-HQ pretrained LDM, increasing guidance strength δ from 0 to 4–5 reduces gender fairness disparity (FD) from 0.051 to 0.001, about a 98% cut, while FID moves from 34.66 through a best region near 30.17 before rising again at overly strong δ. The abstract and body also report about 95% FD reduction for binary race and 15% for four-class race, with perceptual quality held across groups. Training the boundary on 15,000 synthetic samples already reaches FD 0.001 and FID 31.54; 45,000 samples barely change those figures. Multi-attribute tables show joint gender–race balancing raises FD relative to single-attribute runs, as expected when more constraints share one latent nudge. 13 14

Code/demo. The retrieved arXiv record and HTML paper list no public code, demo, or project URL. 13

Caveat. SBP still needs an attribute classifier to label the synthetic boundary-training set. Four-class race remains much harder than binary attributes. Over-large δ trades fairness for FID, and the method is demonstrated on face LDMs rather than general text-to-image stacks. 14

Why it matters. Fairness work on diffusion often forces a retrain or a long guided sampler. SBP is a compact latent-space claim: demographic geometry is sharper late, editable early, and one orthogonal nudge can move mass without touching weights. Read it if you audit synthetic face data or design inference-time edits that must stay cheap.

Where to start

These five papers put diffusion in five different jobs on the same day:

Interactive 4D generator: 4DStreamCtrl unifies camera, object, and depth tracks and still streams at ~20 FPS.
Reference-faithful restorer: GraftSR grafts identical-instance texture through dual masks instead of hoping the prior invents the right surface.
Representation auditor: DEFUSE turns a conditional diffusion prior into an SSL backdoor score, with ACM MM acceptance and public code.
Streaming motion controller: InteractGesture lets future joint targets rewrite earlier gesture chunks without retraining.
One-shot fairness editor: SBP learns demographic boundaries late and applies them once at the noisy start.

Choose the original by the bottleneck you are studying. Start with 4DStreamCtrl for online geometric video control, GraftSR for identity-preserving generative SR, DEFUSE for encoder security with generative priors, InteractGesture for chunked test-time motion guidance, and SBP for cheap demographic rebalancing of face LDMs. Each result is a fresh arXiv preprint, so the reported numbers are author results awaiting the scrutiny a full read provides.