Windflow FAQ: realtime video and world models
Answers about Windflow video and world model APIs: supported inputs, live sessions, per-second pricing, private-beta access and on-demand capacity.
Short answers, written to be quoted. Each links to the page with the detail.
FAQ
What is Windflow?
Windflow is a developer platform that runs realtime video and world models as on-demand sessions. Developers plug a model into their product through one API and pay by the second. It is built by Windflow Labs and is in private beta. See About.
What is a realtime model?
A realtime model generates output continuously while you steer it, instead of returning one finished file after a wait. Examples are a video model that keeps generating as the prompt changes, a world model you walk through with WASD controls, and a model that edits a live webcam feed. The glossary defines the terms.
Which models does Windflow host?
12 public models: Helios, LongLive-2.0, SANA-Streaming, EdenLive, JoyAI-Video-Edit, LingBot-World 2.0, Matrix-Game 4.0, DreamX-World-5B, ABot-World-0-5B, MiniMax FastH3, LTX-2.5, Hy-Nomas. They cover continuous long video, multi-shot video, live video editing, navigable worlds, neural game rendering, and video with generated audio. Each has its own page in the model catalog.
How does a session work?
You create a session for a model, connect a media transport, send controls such as prompts, images, camera motion or actions while the model keeps running, and receive the live stream on a named output track. You release the session when you are done. Control goes over REST; media goes over WebRTC or WebSocket. See the platform page.
How is Windflow billed?
By the billable session-second. Billing starts at the session's first generation_started event and stops when the session is released or ends. Queueing, worker startup, time before generation starts and the documented reconnect grace are not billed. After generation starts, paused or unwatched time remains billable; release the owner session to stop the meter. There is no subscription, no seat fee and no minimum. See pricing.
How much does it cost?
Rates run from $0.0017 to $0.0069 per billable second depending on the model, which is $6.12 to $25.00 per billable hour. Most models sit between $0.0017 and $0.0024 per second ($6.12 to $8.64 per hour). A step-by-step estimate is in how per-second pricing works.
Is there a free trial?
Approved accounts start with $5 of usable balance and four concurrent sessions. That is enough to run several short sessions across the public models.
How do I get access?
Windflow is in private beta. Request access or email hello@windflow.online; a live demo of any model can be arranged on request.
Are the GPUs always running?
No. GPU capacity is brought up on demand rather than kept standing. Live readiness for each model is published at https://api.windflow.online/v1/status, and a live demo can be arranged on request.
Which GPUs does Windflow use?
NVIDIA and AMD GPUs. Workers are provider-neutral and are placed where capacity and cost fit the model.
Which video transports are supported?
WebRTC and WebSocket (H.264). A model negotiates the transports it supports, in order of preference. For the trade-offs between them, see streaming video diffusion over WebRTC or WebSocket.
Can I send a webcam feed?
SANA-Streaming, EdenLive, JoyAI-Video-Edit accept live source video in their published Windflow profiles. Choose by the model's supported controls and input contract. See live video editing from a webcam.
Can I host my own model on Windflow?
Yes. Windflow hosts and optimises a realtime model on NVIDIA and AMD GPUs and wraps it in the same session API. There are three ways to work together; see bring a model.
Does pausing the video stop the meter?
No. Pausing a track only stops delivery to that connection, and a generation pause can still leave the GPU reserved. Release the session when the experience is finished or idle.
Do extra viewers cost more?
No. One session is metered once, however many authorised guests receive it. Separate concurrent sessions each carry their own meter.
Which licences apply to the models?
Each model keeps the licence its authors published. Every model page links to its upstream repository and states the licence. FastH3 and LTX-2.5 are covered by the MiniMax H3 Community License and the LTX-2 Community License respectively; read them before building a product on those models.
Is every model in the world model directory available through Windflow?
No. The directory is a research dataset covering sources and providers beyond Windflow. The hosted catalog describes Windflow's own public service profiles. Follow a model's specific API guide and current access requirements before planning an integration.
Does a model page marked production guarantee an immediately available GPU?
No. Implementation status describes a service profile's maturity. It is separate from private-beta approval and current readiness. A listed model, a ready worker and a session with a visible first frame are different states; check each rather than treating a public page as a live service test.
Where can I find Windflow's original performance data?
The SANA camera report, JoyAI latency report and EdenLive compute report publish selected JSON/CSV results and their measurement boundaries. They describe dated experiments, not a platform-wide latency or quality ranking.
Does a higher displayed FPS mean a more responsive model?
Not necessarily. Interpolation can raise displayed cadence, and queues can make a smooth picture show old input. Record native frame rate, presented frame rate, freezes and control-to-visible delay separately. The measurement guide explains those boundaries.
Does an edit instruction guarantee that motion, identity and the background stay unchanged?
No. The instruction describes the intended change, while the model generates the output. Evaluate source/output pairs, full clips and transitions on your own inputs. Claims about exact preservation require evidence for that specific model, configuration and use case.
Where is the API documentation?
The full API reference is served to Windflow accounts. A public plain-text overview for people and language models is at /llms.txt, and current public rates are available at GET /v1/pricing.
Is Windflow the same as windflow.io or windflow.ai?
No. Windflow Labs at windflow.online is not affiliated with the AI testing product at windflow.io, the enterprise AI company at windflow.ai, the WindFlow C++ library or the wind-turbine company Windflow Technology. See About.