Video Understanding

VBench 2.0: A new benchmark suite advancing video generation evaluation with intrinsic faithfulness metrics.

This survey explores the evolution of physics cognition in video generation, addressing the gap between visual realism and physical accuracy.

Synthetic data can enhance the physical realism of video synthesis, paving the way for more believable generated content.

Opt-CWM: Self-supervised motion learning via counterfactual optimization, achieving state-of-the-art without labels!

AccVideo accelerates video diffusion by 8.5x with a synthetic dataset and trajectory-based distillation, maintaining quality and enabling higher resolution video generation.

Video-T1 enhances video generation through test-time scaling, improving quality and consistency by viewing generation as a search for optimal video trajectories.

Hummingbird: An efficient text-to-video model that balances quality and computational efficiency via pruning and visual feedback learning.

OmnimatteZero: Real-time omnimatte using pre-trained video diffusion, no training needed!

4D-Bench: The first benchmark for assessing MLLMs in 4D object understanding, revealing weak temporal understanding and the need for advancements.

FluxFlow: Make your video generator stronger via temporal regularization!

FluxViT: Flexible video models via adaptive token selection for efficient deployment!

MagicComp: Dual-Phase Refinement Enables Training-Free Compositional Video Generation

Impossible videos expose AI limits!

Concat-ID: A universal, scalable framework for identity-preserving video synthesis, balancing consistency and editability.

KDTalker: Accurate & efficient audio-driven talking portrait via implicit keypoint-based spatiotemporal diffusion, unlocking diverse & realistic animations.

MagicID: ID-consistent & dynamic-preserved video customization via hybrid preference optimization.

ReCamMaster: Re-shoots videos via generative rendering, controlling camera movement from a single source, for novel perspectives and enhanced video creation.

MTV-Inpaint: A unified framework for multi-task long video inpainting, enabling versatile object insertion, scene completion, editing, and removal.

LCT: Fine-tunes single-shot video diffusion models for coherent multi-shot video generation without extra parameters!

CINEMA: MLLM-guided coherent multi-subject video generation for consistent and controllable content creation.