Welcome to the Vchitect homepage. Vchitect is mainly developed by Shanghai AI Laboratory. We keep working in the field of video generation, open-sourcing the models, benchmark suites, and efficient training tools.

Vchitect 2.0

• [09/2024] We release Vchitect 2.0, including the model and the training system
  • Model:
    • Vchitect-2.0 is a high-quality video generative model with 2 billion parameters, supporting resolutions up to 720x480 and video durations of 10-20 seconds.
    • VEnhancer is a generative space-time enhancement framework. It integrates super-resolution, frame interpolation, and video refinement to elevate the video quality to 2K resolution at 24 FPS.
  • System:
    • LiteGen is a lightweight and highly efficient training framework for diffusion tasks. It supports sequence lengths of up to 1.63 million tokens using 8x NVIDIA A100 GPU cards during the training of the Vchitect-2.0 model.
    • FasterCache is a training-free method for accelerating video sampling in diffusion transformers.
  • Evaluation:
    • VBench is a comprehensive benchmark suite for video generative models, covering 56 text-to-video generation models.
    • VBench++ further supports image-to-video evaluation, with an Image Suite of high-resolution images and adaptive aspect ratios. It also focuses on trustworthiness of video generative models like fairness, bias, and safety.
    • Evaluation Agent is an efficient evaluation paradigm for visual generative models. It uses dynamic multi-round evaluations, reducing evaluation time to 10% of traditional methods while ensuring efficiency, customization, and explainability.

• 🎉 [new] Vchitect-2.0: A high-quality video generation video with resolutions up to 720x480 and video durations of 10-20 seconds.
• 🎉 [new] VEnhancer: A generative space-time enhancement framework that can improve the existing T2V results.

• 🎉 [new] FasterCache: A training-free method for accelerating video sampling in diffusion transformers.
• 🎉 [new] LiteGen: A light-weight and high-efficient training framework for accelerating diffusion tasks.

• 🎉 [new] Evaluation Agent: An efficient and promptable evaluation paradigm for visual generative models.
• 🎉 [new] VBench++: A comprehensive benchmark suite for video generative models with image-to-video and trustworthiness evaluation support.
• 🎉 [new] VBench: A comprehensive benchmark suite for video generative models

• Latte: Latent Diffusion Transformer for Video Generation

• LaVie: High-Quality Video Generation with Cascaded Latent Diffusion Models
• SEINE: Short-to-Long Video Diffusion Model for Generative Transition and Prediction
• VideoBooth: Diffusion-based Video Generation with Image Prompts
• Vlogger: A generic AI system for generating a minute-level video blog (i.e., vlog) of user descriptions.
• Optix: Memory Efficient Training Framework for Large Video Generation Model