Video-T1 - A video generation technology jointly developed by Tsinghua University and Tencent.
Video-T1 is a video generation technology jointly developed by researchers from Tsinghua University and Tencent. It improves the quality and consistency of generated videos based on Test-Time Scaling (TTS). Traditional video generation models, in training...
What is Video-T1?
Video-T1 is a video generation technology jointly developed by researchers from Tsinghua University and Tencent. It improves the quality and consistency of generated videos based on Test-Time Scaling (TTS). Traditional video generation models generate videos directly after training, while Video-T1 introduces additional computational resources during the testing phase, optimizing video quality by dynamically adjusting the generation path. The research introduces the Tree-of-Frames (ToF) method, dividing video generation into multiple stages to progressively optimize frame coherence and matching with text prompts. Video-T1 provides a new optimization approach for the field of video generation, demonstrating the powerful potential of test-time scaling.
Main functions of Video-T1
- Improve video qualityIncrease computing resources during the testing phase to generate higher quality videos and reduce blur and noise.
- Enhance text consistency: Ensure the generated video matches the given text prompts, improving the matching accuracy between the video and the text.
- Optimize video continuityImproves motion smoothness and temporal coherence between video frames, reducing flicker and jitter.
- Adapt to complex scenariosGenerates more stable and realistic video content when dealing with complex scenes and dynamic objects.
The technical principle of Video-T1
- Search space constructionThe search process is guided by feedback provided by verifiers during test time, combined with heuristic algorithms.
- Random linear searchDuring inference, noisy candidate samples are added, and video clips are generated step by step by removing noise. The result with the highest validator score is selected.
- Tree-of-Frames (ToF) method:
- Image-level alignmentThe generation of the initial frame affects subsequent frames.
- Dynamic prompt application: Dynamically adjust prompts in the test verifier, focusing on motion stability and physical plausibility.
- Overall quality assessmentEvaluate the overall quality of the videos and select the video that best matches the text prompts.
- Autoregressive expansion and pruningThe video branches are dynamically expanded and pruned based on an autoregressive approach to improve generation efficiency.
Video-T1 project address
- Project official website:https://liuff19.github.io/Video-T1/
- GitHub repository:https://github.com/liuff19/Video-T1
- arXiv technical paper:https://arxiv.org/pdf/2503.18942
Application scenarios of Video-T1
- Creative Video ProductionIt enables content creators and the advertising industry to quickly generate high-quality video materials that meet their creative needs, thereby enhancing the appeal of their content.
- Film and television productionIt assists in the production of special effects and animation, generates complex scenes and character movements, and improves the efficiency of film and television production.
- Education and TrainingGenerate instructional videos and training simulations to enhance the fun and intuitiveness of teaching and training.
- Game developmentGenerate game cutscenes and virtual character animations to enhance the game's immersion and interactivity.
- VR and ARGenerate high-quality VR content and AR dynamic effects to enhance user experience and immersion.