Video Diffusion Scaling
Investigates architectural and scale-related improvements for joint video and audio generation with diffusion models.
20 papers
Where this stands
The written synthesis of this thread is for subscribers. Subscribe.
How this thread developed
Oct 2026 · University of Chinese Academy of Sciences, ByteDance
Splitting experts by role improves video diffusion model scaling
Scales video diffusion models using role-based MoE, bifurcating experts for semantic and generic processing to improve generation quality.
2 further papers
Oct 2026 · University of Rochester, Adobe Research
ALIVE inserts objects that respond coherently to video actions
Proposes a diffusion-based video editor that inserts objects responding to actions, adding high-level interaction control to video generation.
4 further papers
Oct 2026 · Tsinghua University, Robert Bosch GmbH
Symmetry-regularized flow matching enables video generation from any camera pose without ground truth novel views
Extends flow matching to multi-view-consistent video generation from arbitrary camera poses without novel-view supervision.
Oct 2026 · Amazon Prime Video, Tel Aviv University
SemCam turns subject-relative camera instructions into adaptive video motion
Formulates semantic camera motion control as a video-to-video task, demonstrating adaptive motion generation with a diffusion model.
Oct 2026 · Google, University of Illinois at Urbana-Champaign
Explicit context constraints improve interactive video story continuations
proposes VIS-Ground framework that explicitly models context dependencies for interactive video story continuation.
Oct 2026 · LIGHTSPEED, Independent Researcher
Parallel adapter composition yields real-time joint audio-video generation without chained training
Combines independently trained adapters for block-causal attention and few-step sampling to enable real-time joint audio-video generation without chained distillation.
2 further papers
Oct 2026 · The Chinese University of Hong Kong, Shenzhen, Tencent Hunyuan
Causal video diffusion matches bidirectional quality without a teacher model
Proposes Contitional Residua Predction to enable caual video diffusion matching bidirectional quality without a teacher.
2 further papers
Oct 2026 · Renmin University of China, WeChat, Tencent Inc.
Human-rated clips improve rewards for synchronizing generated sound and video
creates a benchmark and reward model for audio-video synchronization, improving joint generation quality.
Oct 2026 · Yonsei University, Microsoft Research Asia
Explicit predictive states boost physical plausibility in video world models
introduces a new architectural component (Physical Registry) for video diffusion models to encode dynamics.
Oct 2026 · SenseTime, Beihang University
LightVAE speeds up video autoencoders by exploiting temporal and depth redundancy
Accelerates video autoencoder decoding with temporal low-rank compression and width-priority distillation, reducing a bottleneck in video generation pipelines.
10 of 20 papers shown