Pyramid Flow (Pyramidal Flow Matching for Efficient Video Generative Modeling) is a model aimed at efficient video generative modeling, whose core method is pyramidal flow matching. According to the official page information, this model is a training-efficient autoregressive video generation model built on Flow Matching.
In terms of use: Pyramid Flow is mainly used for video generation tasks, specifically including Text-to-Video Generation and Text-conditioned Image-to-Video Generation.
Main capabilities and results: The official page presents qualitative results, including:
- Text-to-video generation, resolution 1280x768, duration 10 seconds, frame rate 24fps;
- Text-to-video generation, resolution 1280x768, duration 5 seconds, frame rate 24fps;
- Text-conditioned image-to-video generation, resolution 1280x768, duration 5 seconds, frame rate 24fps. In addition, the official page also provides quantitative results, involving VBench (Huang et al., 2024) and User Preference.
How to use: The official page does not provide specific installation commands, code examples, or API usage instructions; none are currently provided.
Training details: The official page states that the model is trained using only open-source datasets, and the training consumption is 20.7k A100 GPU hours.
Applicable scenarios: Based on the capabilities it demonstrates, it is suitable for text-to-video and image-to-video generation scenarios, especially research and applications focused on training efficiency and generation metrics such as resolution, duration, and frame rate.
Limitations: The official page does not clearly list model limitations, license agreements, commercial use policies, or pricing information; these are currently not provided.
