Vidu S: a repository that points to a hosted real-time video API

What Vidu S actually ships in the repository

The repository at shengshu-ai/Vidu-S is not a model release. Its top-level entries are README.md and figures/, and the README is a landing page for a hosted service rather than a codebase. There is no training script, no inference server, no checkpoint link and no licence file in what the repository exposes. The README links to a paper for Vidu S2 (arXiv:2609.11638) and one for Vidu S1 (arXiv:2607.03118), plus English and Chinese user guides on Feishu and API documentation on platform.vidu.com and platform.vidu.cn. The last push was on 2026-09-16, and the repository is not archived, so the README is current. But currency of a README is not the same as a maintained library. Anyone who clones this expecting to build and serve a video model will find nothing to compile. The practical entry point is the hosted API and the Vidu S2 demo at vidu.com/vidu-stream.

The problem Vidu S targets: streaming generation, not clip rendering

Most text-to-video tooling is built around a prompt-and-wait cycle. You submit a prompt, wait for a clip, inspect it, and resubmit. Vidu S is aimed at the opposite shape of workload: a continuous stream where the user changes something mid-session and the output responds without restarting. The README describes three such workloads. Vidu S2-Avatar generates a controllable character that accepts new reference images at any moment during a stream, including large body motions such as dancing. Vidu S2-Editing transforms an incoming video stream with style transfer, virtual try-on, character replacement or background replacement from text instructions, while preserving the source motion. The spatial mode converts a generated or edited stream into synchronized left- and right-eye views, and can edit stereoscopic input jointly for VR headsets. This is a fit for live avatar presenters, interactive try-on, and headset streaming. It is a poor fit for offline batch rendering of a fixed shot list, where a conventional diffusion pipeline is simpler and cheaper.

Self-Replay Forcing and the serving stack behind the frame rate

The README names two mechanisms. The first is Self-Replay Forcing (SRF), described as replaying re-noised, self-generated trajectories in a gradient-enabled causal pass so that errors do not accumulate across streaming segments. That is the answer to the standard failure of autoregressive video: drift. Each generated segment becomes the conditioning context for the next, and small errors compound into visible degradation over long horizons. SRF is the stated countermeasure. The second is the serving stack, which combines TurboDiffusion and TurboServe and is described as using efficient attention, low-bit GEMM, kernel and launch optimizations, and multi-GPU pipelining. The performance claims attached to these are 720p at 25 to 42 FPS for Vidu S2-Avatar and 540p at up to 42 FPS for Vidu S1, with S1 stated to run on consumer GPUs. The README does not say which GPU models, batch sizes or resolutions those numbers assume, and it does not publish latency percentiles. Treat the frame-rate figures as vendor targets until you reproduce them on your own stream.

Installing the agent skill and making a first API call

There is no local install. The README's only installable artifact is an agent Skill named vidu-s-api, hosted in a separate repository at shengshu-ai/vidu-s-api under skills/vidu-s-api. In Claude Code, Codex, OpenClaw or another agent that supports Skills, the README says to paste this instruction:

Install this skill: https://github.com/shengshu-ai/vidu-s-api/tree/main/skills/vidu-s-api

The agent clones the skill into its Skills directory. The README notes that you may need to restart the agent, after which you ask it to load vidu-s-api and it will walk through Vidu S2 API integration. That is the whole setup path documented in this repository. For a direct integration without an agent, the README points to the API documentation rather than reproducing request examples, so the request shape, authentication header, endpoint host and streaming transport must be read from platform.vidu.com/vidu-stream/doc. The two quickstarts worth opening first are the S2-Avatar realtime page and the S2-Editing page, since those correspond to the two API surfaces the README describes in most detail. If you are evaluating rather than integrating, the README's Try Vidu S2 link goes to the hosted demo, which requires no repository setup at all.

Where Vidu S is the wrong tool

The clearest limitation is distribution. Nothing in this repository lets you run the model yourself, audit it, or pin a version. You are depending on a hosted endpoint whose uptime, rate limits, regional availability and pricing are outside the repository's control, and the README does not document rollback, version pinning or a deprecation policy for the API. A second limitation is the state of the artifact. There is no licence file in the repository, so the terms that govern the code and any weights are not stated here; the README's citation block asks researchers to cite the papers, which is a request, not a licence grant. Third, the frame-rate claims are conditional. Real-time interactive generation at 720p depends on multi-GPU pipelining, so a single-GPU deployment is not what the numbers describe. If your requirement is on-premise inference, reproducible checkpoints, or a licence you can read before shipping, this repository does not answer any of those, and the API docs are the only place to look.

How Vidu S differs from open video generation stacks

The natural comparison is an open video generation project you can self-host, such as a diffusion model served behind your own inference server. The difference is not quality, it is where the boundary sits. A self-hosted stack gives you the weights, the scheduler, the VRAM budget and the failure modes, and you own the latency curve. Vidu S gives you an HTTP surface, a documented frame-rate target, and no visibility into the serving internals beyond the names TurboDiffusion and TurboServe. The trade is operational burden against control. If your workload is bursty and you would rather not hold GPUs, the hosted route removes capacity planning. If your workload is continuous and latency-sensitive, the hosted route adds a network hop you cannot tune, and the README does not publish a latency budget for it. Note also that Vidu S1 and S2 are not interchangeable: S1 is voice-controlled digital characters at 540p, S2 adds editing and spatial output at 720p, so a comparison against a single open model should be scoped to whichever tier you actually need.

Maintenance, upgrade cost and licence status

The repository was last pushed on 2026-09-16, one day before this writing, and it is not archived. That tells you the README is being kept current, not that a library is receiving fixes; there is no library. Upgrade cost therefore lives on the API side. The README's Updates section lists two entries: Vidu S2 became available to try in 2026-09, and Vidu S1 became available in 2026-07. There is no changelog for the API, no versioned endpoint list, and no migration note for anyone who integrated against S1 and now wants S2-Avatar or S2-Editing. Budget for reading the platform documentation at each release rather than for bumping a dependency. On licensing, the repository states no licence, so the code and any associated weights have no terms granted here. The papers are cited under arXiv preprints. Whether commercial use of the hosted API is permitted, and under what terms, is a question for the platform documentation and the vendor, not for this repository. This is a description of what the repository does and does not state, not legal advice.

Editorial conclusion

Adopt Vidu S if you need hosted 720p interactive avatar or live editing streams and can accept a closed, API-only dependency. Do not adopt it if you need to run inference on your own hardware or inspect training and serving code, because the repository contains only a README and a figures directory. Before committing, open the S2-Avatar realtime quickstart at platform.vidu.com/vidu-stream/doc/s2-avatar/realtime/quick-start and confirm the streaming transport, the reference-image swap semantics during a live session, and whether the 25 to 42 FPS range is guaranteed or best-effort on the shared endpoint.

Frequently asked questions

Official sources

1. Issues
2. Project website
3. README
4. shengshu-ai/Vidu-S on GitHub