Seaweed-7B: Cost-Effective Training of Video Generation Foundation Model
- Document
- 11 April 2025
- Event
- 11 April 2025
- Retrieved
- 16 September 2026
The shot
On 11 April 2025, researchers from ByteDance's Seed research group posted a technical report describing Seaweed-7B, a video-generation foundation model built as a diffusion transformer with roughly seven billion parameters. The paper states the model was trained from scratch using 665,000 H100 GPU hours, a deliberately moderate budget compared with larger contemporary systems. A companion page, Seaweed's own project site, frames the same work as "a research effort for building a foundational model for video generation," short for "Seed-Video," and credits ByteDance's Seed leadership in its acknowledgements.
What the documents show
The paper's claim that Seaweed-7B "achieves performance comparable to, or even surpasses" larger video models is ByteDance's own self-reported benchmark result, not an outcome an outside evaluator has confirmed. The project page adds detail the paper does not: it shows the model paired with a list of separately named research efforts built on top of it, including a reference-conditioning system called Phantom, a multi-shot storytelling method called Long-Context-Tuning, a real-time variant called Seaweed-APT, a camera-control system called CameraCtrl-II, and a physical-consistency post-training method called SimDrop. Each is presented as its own project rather than a single finished feature set, which matters for reading the page's demonstration reels: a capability shown once, credited to a named side project, is not the same as a capability built into one downloadable model.
The workflow
Neither document offers a license, a download link, or pricing: the project page directs visitors toward the research paper rather than toward an API or a weight repository, consistent with the paper's own framing of Seaweed as a research system rather than a released product. There is no pipeline step for a production to adopt today. The workflow fact the record actually establishes is more limited: it shows how a major platform company describes its own training economics and architecture, which is useful for tracking research direction, not for planning a shoot.
What the tool does not change
A demonstration reel of generated short films, landscapes, and human-centric video, however extensive, does not establish general production reliability; this site's distinction between a verified production use and a promotional demonstration applies directly here, since every example on the project page is ByteDance's own selection. A self-reported comparison against "larger models" is also not the same claim as a third-party benchmark result. This is an editorial caution, not a statement the sources make themselves.
- Has ByteDance published a license or access path for Seaweed-7B beyond the research paper and demo page?
- Which named capability in the demonstration reel belongs to a separate research project rather than the base model itself?
- What would an independent, non-vendor-run comparison need to show before a production treated the paper's benchmark claim as confirmed?
Seaweed-7B is best read as a data point in ByteDance's research trajectory rather than a new tool available to a production today, and the distance between the two is exactly what its own publication record shows.
Sources & reading trail
States the model's 7B parameter count, training compute of 665,000 H100 GPU hours, and the paper's own comparative performance claim.
Source published: 11 April 2025 · Retrieved: 16 September 2026
ByteDance Seed's own description of Seaweed as a research effort and its list of separately named downstream capabilities (Phantom, Long-Context-Tuning, Seaweed-APT, CameraCtrl-II, SimDrop).
Source published: Not established · Retrieved: 16 September 2026
Documentation, agreements and rulings establish the note; the workflow reading is Screen Method editorial analysis. This retrospective draft does not imply the site published on the event date.