Choosing a realtime video model: a decision guide for 12 hosted models
Pick between 12 hosted realtime video and world models by inputs, live controls, audio, transports, runtime status and current API prices.
Choose a realtime model by the inputs your product already has and the output it must produce, then check transport, licence and price. Windflow hosts 12, and the first question usually settles it: do you need a world to walk through, a video that keeps going, an edit of a live camera, or audio with the picture?
Start with the job
| If your product needs to... | Look at | Why |
|---|---|---|
| Generate continuous video you steer with text | Helios, LongLive-2.0 | Both keep generating and accept prompt changes while running |
| Direct a sequence of shots | LongLive-2.0 | One text prompt per shot, soft changes or scene cuts |
| Generate video with synchronised sound | LTX-2.5, MiniMax FastH3 | The only two that generate audio with the picture |
| Let a user move through a generated place | LingBot-World 2.0, DreamX-World-5B, ABot-World-0-5B | Persistent movement and look controls from a seed image |
| Let a user play a game that a model draws | Matrix-Game 4.0 | A real game runs the rules; a scene model redraws every frame of three bundled games |
| Restyle game footage using source motion | EdenLive | Proxy-guided rendering from a source clip or live track |
| Change a live webcam or a clip with a sentence | SANA-Streaming, JoyAI-Video-Edit | Source video in, edited stream out, motion preserved |
All 12 at a glance
| Model | Inputs | Delivered output | Transports | Price per billable second | Status |
|---|---|---|---|---|---|
| Helios | Text prompt (required before start) and an optional reference image for image-to-video | Negotiated per session | WebSocket, WebRTC | $0.0017 | production |
| LongLive-2.0 | Text shot prompts only — no reference image in this release | 1280 × 704 | WebSocket, WebRTC | $0.0021 | production |
| SANA-Streaming | A source clip you upload (≥ 41 frames) or a live webcam track, plus an edit prompt | 1280 × 704 native edit and delivery | WebSocket, WebRTC | $0.0019 | production |
| EdenLive | Source game footage or a live video track, a scene prompt, and an optional visual reference | 1280 × 704 | WebSocket, WebRTC | $0.0021 | preview |
| JoyAI-Video-Edit | A source clip you upload or a live webcam track, an edit instruction, and optionally a reference image | 840 × 480 interactive tier; 1248 × 720 detail tier | WebSocket, WebRTC | $0.0021 | production |
| LingBot-World 2.0 | A seed image (uploadFile) and a text prompt, plus persistent WASD movement and look controls | 1408 × 768 delivery (704 × 384 generation) | WebSocket, WebRTC | $0.0024 | production |
| Matrix-Game 4.0 | A scene choice and held keyboard controls; no prompt or seed image | 864 × 480 | WebSocket | $0.0024 | preview |
| DreamX-World-5B | A seed image, a text prompt, and movement/camera-control state | 704 × 1280 portrait in the production runtime; negotiated per session | WebSocket, WebRTC | $0.0023 | production |
| ABot-World-0-5B | A seed image, a text prompt, and persistent movement/camera key state | 1344 × 768 delivered (896 × 512 generation + learned super-resolution) | WebSocket | $0.0023 | production |
| MiniMax FastH3 | A text prompt with visual, dialogue, soundscape, and music direction | 960 × 544 | WebRTC | $0.0069 | production |
| LTX-2.5 | A text prompt and an optional reference still | 832 × 448 measured; the canvas is configurable in multiples of 64 | WebRTC | $0.0040 | production |
| Hy-Nomas | Text and persistent camera movement/look snapshots | 832 × 480 | WebSocket, WebRTC | $0.0047 | preview |
Resolutions are what the catalog publishes; Helios negotiates its size per session. The pricing guide turns these rates into hourly and ten-minute costs.
Decision rules
- You need audio. LTX-2.5 or MiniMax FastH3. Both are WebRTC-only.
- You have a webcam feed. SANA-Streaming or JoyAI-Video-Edit. Choose JoyAI-Video-Edit for a reference image or the 720p detail tier; SANA-Streaming is prompt-only.
- You want a world with a text prompt. LingBot-World 2.0, DreamX-World-5B or ABot-World-0-5B.
- You want something to play. Matrix-Game 4.0 runs a castle shooter, an Ancient Egypt shooter or city racing and redraws every frame, in 60-second sessions. It renders only those three games.
- Your audience is on phones. DreamX-World-5B delivers portrait.
- You want a third-person character. ABot-World-0-5B. It is WebSocket-only.
- You want a low-cost way to start. Helios needs only a prompt and, at $0.0017 per second, has the lowest rate in the Windflow catalog.
- You need a stable production model. 9 models are in production. Current previews: EdenLive, Matrix-Game 4.0, Hy-Nomas.
- Licence matters to your product. FastH3 is governed by the MiniMax H3 Community License and LTX-2.5 by the LTX-2 Community License; read them first. Most of the others publish their code under Apache-2.0 or MIT; each model page links to its upstream source.
Mistakes to avoid
- Choosing on price alone. A cheaper model that lacks the input you have, such as a webcam track or a seed image, costs more in rework than it saves.
- Ignoring transport. If your client only supports WebRTC, ABot-World is out; if it only supports WebSocket, LTX-2.5 and FastH3 are out.
- Expecting a game engine. World models regenerate the view; they do not keep exact geometry or rules. See navigable world models for games.
- Assuming the GPU is waiting for you. Capacity is brought up on demand, so plan for a queue and a first-frame delay and test on the model you ship.
A practical way to decide
- Write down the inputs your product already collects.
- Cross out models that cannot use them.
- From what remains, run a short real session on two candidates with the same prompt or seed.
- Compare the control feel, not just the picture, and price the winner for your expected session length.
Match evidence to the decision
For a webcam product, inspect the source-to-visible path. The SANA camera experiment illustrates why approximately 25 presented fps can coexist with approximately 2.8 seconds of scene-change delay in a specific configuration. JoyAI's recorded latency stages separate frame processing from socket emission in a loopback probe; those results cannot establish a browser latency by themselves.
For game-footage transformation, the EdenLive compute report can help select a GPU/software configuration to evaluate. It excludes encoding and client display, so a passing model-loop FPS is the start of a delivery evaluation.
Also distinguish code licenses, weight licenses and hosted-service access. The world model directory records declarations and links primary sources; check the exact model version and dependencies for your intended use. Being able to request an API session does not resolve the rights required by a commercial application.
New accounts start with $5, enough to try several models. Request access.
FAQ
Which realtime video model should I use?
Pick by inputs and output: Helios or LongLive-2.0 for steerable continuous video, LTX-2.5 or MiniMax FastH3 when you need generated audio, SANA-Streaming or JoyAI-Video-Edit for webcam editing, and LingBot-World, Matrix-Game, DreamX-World or ABot-World for navigable worlds.
Which model is cheapest?
In the Windflow catalog, Helios, at $0.0017 per billable second ($6.12 per hour of billable time). SANA-Streaming is $0.0019, and 9 models are under $0.0025.
Which models support webcam input?
SANA-Streaming, EdenLive, JoyAI-Video-Edit.
Which models generate audio?
LTX-2.5 and MiniMax FastH3. Other catalog models are video-only.
Which model runs in portrait?
DreamX-World-5B delivers 704 × 1280 portrait video.
Can I switch models without rewriting my integration?
Yes. Every model speaks the same typed session protocol, so switching changes the controls you send, not how you integrate.