Seaweed by Bytedance

Summary

Seaweed is a video generation foundation model developed by ByteDance, designed to deliver high-quality, realistic videos from text prompts, uploaded images or audio cues. Short for Seed-Video, it is a research effort built on a highly efficient 7-billion-parameter diffusion transformer trained using compute equivalent to 1,000 H100 GPUs, and it stands out for its ability to generate consistent, multi-shot, long-form stories with remarkable semantic understanding and visual fidelity.

Key Features

7-billion-parameter video generation foundation model from ByteDance
Real-time video generation at 1280x720 resolution and 24fps
Text-to-video, image-to-video, reference-based and audio-conditioned generation
Consistent multi-shot and long-form storytelling with global and shot-level control
Native shots up to 20 seconds, extendable to around a minute
Dynamic cinematography with varied camera movements and expressive visuals
Supports a wide range of artistic styles and multiple aspect ratios
Efficient training using 665,000 H100 GPU hours with performance comparable to much larger models

About Seaweed by Bytedance


A hallmark of Seaweed is its dynamic and engaging cinematography: the model can interpret intricate prompts to create vibrant actions, expressive character interactions and varied camera movements such as zooming, panning and tracking, resulting in videos that look visually stunning and feel professionally directed. It supports real-time video generation at 1280x720 resolution and 24 frames per second, native shots up to 20 seconds that can be extended to around a minute, and multiple aspect ratios including 1:1, 16:9 and 9:16, across a broad spectrum of artistic styles from photorealistic animation to stylized 2D and 3D renderings.


Seaweed is engineered for efficiency: thanks to its compact model size and optimized training strategy, it achieves performance that rivals or surpasses much larger models while requiring significantly less computational power, ranking among the top models in text-to-video benchmarks. It is presented as a research showcase rather than a commercial product, with no public pricing or commercial API, and it is suited to researchers and creators exploring video generation, long-form storytelling and real-time interactive applications.

Get more likes & reach the top of search results by adding this button on your site!

Embed button preview - Light theme
Embed button preview - Dark theme
TurboType Banner

Subscribe to the AI Search Newsletter

Get top updates in AI to your inbox every weekend. It's free!