Live video editing from a webcam with a text instruction
Edit a live webcam or uploaded clip using text instructions. Compare SANA-Streaming and JoyAI-Video-Edit inputs, reference conditioning and responsiveness.
Real-time AI video editing takes a live webcam track or an uploaded clip, applies an edit described in text, and returns the edited stream with the original motion preserved. Windflow hosts two models for it, SANA-Streaming at $0.0019 per ready second and JoyAI-Video-Edit at $0.0021, behind the same session API.
You send "put him in a red jacket" or "make it look like watercolour" as a prompt, change it while the stream runs, and the model edits only what the instruction names.
The two models
| SANA-Streaming | JoyAI-Video-Edit | |
|---|---|---|
| Size | 2 billion parameters | 16 billion parameters |
| Source | Webcam track or uploaded clip (at least 41 frames) | Webcam track or uploaded clip |
| Instruction | Edit prompt | Edit instruction, optional reference image |
| Output | 1280 × 704 native edit and delivery | 840 × 480 interactive tier; 1248 × 720 detail tier |
| Price per second | $0.0019 | $0.0021 |
| From | NVIDIA's SANA family | JD's open-source group |
Pages: SANA-Streaming and JoyAI-Video-Edit.
Which one to start with
- Start with SANA-Streaming when you want the smaller model, a fixed 1280 × 704 output and the lowest price in the editing pair. It is a good default for live restyling and effects.
- Start with JoyAI-Video-Edit when you want a larger model with a reference-image option for carrying a specific look onto a subject, or when you want the 720p detail tier. SANA-Streaming is prompt-only.
Both preserve the motion of the source, so a speaker keeps moving the way they did while clothing, background or style change. JoyAI-Video-Edit also preserves identity.
What you can build
- Live effects for streamers and video calls: costumes, settings and art styles driven by a text box or a button.
- Prompt-driven filters that are described rather than shipped as fixed presets.
- Clip editing from a text instruction, without a re-render from scratch.
- Virtual try-on style experiences with JoyAI-Video-Edit, where a reference image guides the change.
Design notes
- Edits are scoped. Parts of the source the prompt does not name are preserved. Write instructions that name what should change.
- JoyAI-Video-Edit holds the picture briefly on a prompt change. A new instruction, or a new reference image, reopens the model's session at the next boundary. Design the interface so that beat does not read as a freeze.
- SANA-Streaming re-anchors periodically to limit drift in long live runs, and a refresh frame can be visually abrupt. Listen for the anchor event and mask it if you need a seamless stream.
- Fix the camera resolution before you start a SANA-Streaming live session; a resolution change mid-stream can destabilise it.
- Consent and disclosure. Editing a person's face or clothing in live video is sensitive. Tell users when the stream is AI-edited and get consent from anyone on camera.
Getting started
- Request access and choose a model.
- Mint a short-lived token on your server and open a session from the client.
- Publish the camera track, or upload a clip, and send the first edit prompt.
- Change the prompt as the user asks; release the session when they leave.
The general flow is in add live AI video to your app. A ten-minute SANA-Streaming session costs $1.14 and a ten-minute JoyAI-Video-Edit session costs $1.26.
FAQ
Is there an API for real-time AI video editing from a webcam?
Yes. Windflow hosts SANA-Streaming and JoyAI-Video-Edit, which accept a live webcam track or an uploaded clip plus a text edit instruction and return the edited stream in real time, billed per ready second.
Does it edit the whole frame?
No. Edits are scoped to what the instruction names; unmentioned parts of the source are preserved, including its motion. JoyAI-Video-Edit also preserves identity.
Can I use a reference image?
With JoyAI-Video-Edit, yes: an optional reference image helps carry a specific look onto the subject. SANA-Streaming is driven by the edit prompt alone.
What resolution do I get?
SANA-Streaming edits and delivers at 1280 × 704. JoyAI-Video-Edit offers an 840 × 480 interactive tier and a 1248 × 720 detail tier, and super-resolution can be applied to the delivered stream.
Does it work on a live camera, or only on files?
Both. Live mode edits a camera track as it arrives; file mode edits an uploaded clip.