Five diffusion papers: July 9-10, 2026

This two-day catch-up covers the arXiv window from July 8, 09:00, to July 10, 09:00 UTC-5. Because there was no July 9 issue, the digest selects the strongest five papers from the full catch-up window rather than forcing a normal one-day cut.

The selected set is unusually concentrated: all five papers sit in video, relighting, event-based reconstruction, or human motion. That means diffusion-language and theory papers with good reported numbers are left out today, while the ranked list follows the window's strongest video/motion cluster.

Decision: open LightCrafter first if your work touches video relighting, physically based rendering conditions, or consistency constraints in controllable video generation. The paper ranks first because it connects a high-value production problem, relighting, to video diffusion refinement rather than treating illumination as a cosmetic post-process.

Paper signal: LightCrafter: PBR-Conditioned Video Diffusion Refinement for Controllable and Consistent Relighting is listed as a CMU source in the July 8-10 diffusion window. 1 PBR means physically based rendering, so the title points to conditioning signals that should carry lighting, material, and scene-geometry information more explicitly than a text prompt alone.

Technical read: the interesting claim is the word "refinement." A practical relighting system has to preserve identity, geometry, motion, and temporal coherence while changing illumination. If LightCrafter uses PBR conditions as a structured control channel, the full paper should clarify whether the diffusion model is correcting rendered estimates, harmonizing them across time, or using them as a stronger conditioning prior during video generation.

Evidence and gaps: Metrics: no concrete benchmark values were available. Code/demo: not verified. That makes the first read a method-inspection read: check the comparison set, temporal consistency metrics, human preference setup, and whether relighting control is tested beyond curated indoor or object-centric scenes.

Read it for: whether PBR-conditioned refinement gives a reusable control interface for video diffusion. If the method handles moving cameras, changing materials, and long clips without identity drift, it belongs near the top of the relighting stack.

Decision: open Gen4U if you care about representation learning inside video diffusion models. It ranks second because the question is broader than sample quality: can one diffusion formulation support both generating video and understanding video?

Paper signal: Gen4U: Unifying Video Generation and Understanding via Diffusion is listed as a Google DeepMind source in the catch-up window. 2 The Google DeepMind affiliation signal remains provisional because the author affiliation was not fully verified.

Technical read: the full paper should be read for the interface between generative training and discriminative or semantic video tasks. A useful unification paper has to do more than attach a classifier to a generator. It should show how diffusion-time representations, latent trajectories, or denoising objectives transfer into understanding tasks without destroying generation quality.

Evidence and gaps: Metrics: no concrete benchmark values were available. Code/demo: not verified. The first read should focus on task coverage: which video-understanding benchmarks are used, whether generation and understanding are evaluated on the same backbone, and whether the paper compares against video foundation models that were not trained as diffusion generators.

Read it for: the bridge between video diffusion as a sampler and video diffusion as a representation learner. If Gen4U makes that bridge clean, it could matter for labs that want one model family to support synthesis, retrieval, captioning, and downstream video reasoning.

Decision: open OPSD-V if sampling cost is your bottleneck in autoregressive video generation. Among today's five, it has the clearest surfaced quantitative signal.

Paper signal: OPSD-V: On-Policy Self-Distillation for Post-Training Few-Step Autoregressive Video Generators is listed as an HKUST source. 3 The title places the paper in a specific post-training lane: few-step generation after an autoregressive video model already exists.

Technical read: the phrase "on-policy" is the part to inspect. In few-step distillation, the student can drift away from the teacher distribution because its own short trajectories differ from the teacher's longer sampling process. An on-policy setup should expose the student to its own generation states during post-training, which may reduce mismatch between training and inference.

Evidence and gaps: Metrics: OPSD-V reports 66.0% overall user-study preference and 82.5% preference when ties are excluded. 3 Code/demo: not verified. Those numbers are useful, but they do not replace dataset-level temporal metrics, prompt categories, or ablations.

Read it for: whether on-policy self-distillation is a stable post-training recipe for autoregressive video generators. The full paper should answer what gets distilled, how the policy distribution is sampled, and whether quality holds as clip length increases.

Decision: open LongE2V if you work on event cameras, low-latency sensing, or reconstruction under sparse temporal observations. It is the most domain-specific paper in today's list, but the domain is technically demanding.

Paper signal: LongE2V: Long-Horizon Event-based Video Reconstruction, Prediction, and Frame Interpolation with Video Diffusion Models is listed as an NYCU source and as a SIGGRAPH 2026 paper. 4 Event-camera streams record brightness changes rather than dense RGB frames, so reconstruction and interpolation require a model to infer missing visual state from asynchronous signals.

Technical read: the title bundles three tasks: reconstruction, prediction, and interpolation. That combination matters because event-video systems often need all three in sequence. A strong paper should show whether one video diffusion model handles the long-horizon setting directly or whether the method uses task-specific heads, conditioning paths, or temporal stitching.

Evidence and gaps: Metrics: no concrete benchmark values were available. Code/demo: not verified. The full paper should be checked for dataset choice, horizon length, latency assumptions, and whether SIGGRAPH-level visual results are supported by quantitative comparisons against event-based reconstruction baselines.

Read it for: the conditioning design. If LongE2V can keep long-horizon predictions stable from event streams, it may give video diffusion researchers a cleaner way to work with sensors that do not produce conventional frames.

Decision: open ARDY if your work involves human motion generation, interactive control, or hybrid continuous/discrete representations. It ranks fifth because the venue and lab signals are strong, but the surfaced evidence is thin.

Paper signal: ARDY: Autoregressive Diffusion with Hybrid Representation for Interactive Human Motion Generation is listed as an NVIDIA source and as a SIGGRAPH 2026 paper. 5 The title points to two design choices: autoregressive rollout and a hybrid representation for motion.

Technical read: human motion generation is sensitive to representation. Joint positions, rotations, contacts, trajectories, and high-level actions do not fail in the same way. A hybrid representation may let the model keep interactive controllability while preserving motion realism, but the full paper needs to show which variables are generated autoregressively and which are handled through diffusion refinement.

Evidence and gaps: Metrics: no concrete benchmark values were available. Code/demo: not verified. For a full read, check latency, control interface, contact quality, and whether the method is evaluated under interactive edits rather than only offline motion synthesis.

Read it for: the representation choice. If ARDY gives a clean interface between user control and diffusion-based motion synthesis, it could be useful beyond character animation, including embodied agents and simulation data generation.

For video control and relighting, start with LightCrafter, then compare its conditioning design against LongE2V's event-stream conditioning. 1 4

For video model scope, read Gen4U before OPSD-V. Gen4U asks whether diffusion can unify generation and understanding, while OPSD-V asks how to make autoregressive video generation cheaper after training. 2 3

For motion and interaction, read ARDY with a narrow question: whether the hybrid representation is specific to human animation, or whether it suggests a broader interface for controllable diffusion rollouts. 5