Runway’s new research Solaris redefines UI creation by treating the interface as a continuously generated visual world, eliminating the need for pre‑written code and enabling real‑time, model‑driven interactions that outperform traditional code‑based approaches in user studies.
DreamWorld introduces a two‑stage Geometry‑then‑Appearance pipeline that first generates geometry features for the target viewpoint using a 3D foundation model and then conditions a video diffusion model on these features to produce high‑quality, 3D‑consistent RGB videos, achieving state‑of‑the‑art results on multiple benchmarks.
LocalDPO introduces a zero‑annotation, region‑aware DPO framework that uses high‑quality real videos as positive samples and automatically generated locally degraded negatives to align video diffusion models with human preferences, achieving significant gains in visual quality, temporal consistency, and subjective ratings on CogVideoX and Wan2.1.
FreeOrbit4D achieves training‑free arbitrary camera redirection for a single monocular video by reconstructing a foreground‑complete 4D geometry, delivering stable large‑angle shots, beating baselines on VBench and user studies, and exposing an editable 4D point cloud for many downstream applications.
CameraSquad introduces a parallel multi‑trajectory video generation framework that delivers precise camera control and cross‑view content consistency, enabling high‑quality 3D point‑cloud reconstruction and superior performance on benchmarks such as WebVid and HumanVid compared with prior camera‑controlled video methods.
AnyFlow introduces a flow‑map distillation framework that enables video diffusion models to produce high‑quality results in just four steps while continuously improving with additional sampling steps, supporting both causal and bidirectional architectures up to 14 B parameters and allowing downstream fine‑tuning.
AnyFlow introduces a flow‑map distillation framework that enables video diffusion models to produce high‑quality results in just four sampling steps while still gaining quality as the number of steps increases, supporting both causal and bidirectional architectures and scaling up to 14 B parameters.
The CVPR 2026 paper introduces NS‑Diff, a physics‑guided video diffusion framework that combines a noise‑robust dynamics detector, a physical‑condition latent injection module, and reinforcement‑learning optimization to reduce jerk error by 43 % and fluid divergence by 33 %, achieving superior physical realism and visual quality across multiple benchmarks.
This article examines the Wan2.1 video diffusion model, identifies its scalability bottlenecks for long and real‑time video generation, and introduces the Self‑Forcing causal framework together with sequence‑parallel and RoPE optimizations that achieve sub‑second latency and up to 1.5× speed‑up on modern GPUs.
FantasyTalking generates high-fidelity, coherent talking portraits from a single static image by employing a two-stage audio-visual alignment—global segment-level motion and frame-level lip refinement—combined with face-centric cross-attention for identity preservation and a motion-intensity module that lets users control expression and body movement, achieving superior realism, synchronization, and performance over prior methods.
This comprehensive review examines Sora, the first model capable of generating minute‑long, high‑quality videos from text, covering its historical background, core diffusion‑Transformer architecture, data preprocessing strategies, prompt engineering techniques, diverse applications, and the ethical and technical limitations that shape its future.
This article presents concise English summaries of nine Kuaishou research papers accepted at ACM MM 2023, covering topics such as no‑reference video quality assessment, adaptive video quality models, blind image super‑resolution, audio‑visual‑language transfer learning, motion‑aware video diffusion, large‑scale e‑commerce retrieval, and interactive segmentation.
