World models and interactive
Decart Lucy 2.6: live video editing with reference images
Decart has released Lucy 2.6, a real-time video editing model that swaps characters, garments and objects in a live stream at 720p. It is the new default behind lucy-latest, at the same $0.02 per second as Lucy 2.5.
On 5 November, Decart released Lucy 2.6, the latest version of its model for editing live video as it streams. The company says it follows the request and the reference image much more closely than Lucy 2.5, and that it is available now on the Decart platform.
What it does
Lucy edits any element of a live video stream: add, replace or remove objects, swap characters, change backgrounds, restyle the scene or apply effects. You can give it a text prompt, a reference image, or both. The docs list 720p output.
Decart names three improvements:
- Characters true to the reference. Faces, proportions and materials follow the reference more closely, for both photorealistic and animated characters, with better lighting, lip sync and motion.
- Clothing that keeps its fit. Complex or unusual garments hold their shape and detail through the stream.
- Items that look like the real thing. Added or replaced props and products follow the reference in detail, material and proportion, and small or off-centre objects are replaced in place at the right scale.
The docs add multi-edit: several operations, such as a background change and a character swap, can be combined in one prompt, although each request still takes one reference image.
Access and price
Lucy 2.6 is the default behind the lucy-latest model name, and Lucy 2.5 stays available as lucy-2.5. Realtime use costs $0.02 per second at standard speed, or $0.04 per second in fast mode, which has lower latency. Decart says the price is unchanged from Lucy 2.5. Video-file editing with the same model is $0.04 per second at 720p. SDKs cover JavaScript, Python, Swift and Android.
Limits
Decart’s examples target live commerce and social streaming: trying on garments, holding products, streaming as a character. The release gives no benchmark numbers, frame rate or latency figures for 2.6, so comparisons with 2.5 rest on the company’s before-and-after clips. Each request takes a single reference image.
Our take
Live, reference-driven editing is a different tool from a world model, but it sits close to one. A stream that can re-dress a person or restage a room in real time is a rendering layer, and game teams can use that for player avatars, mixed-reality sets and live events. It still has no memory or rules of its own. For how the two ideas differ, see how AI world models work.