Decart ships Lucy 2.5: edit video while it streams, latency claimed under 40ms
Add objects, swap backgrounds, run effects live. The product page cites sub-40ms latency and 100FPS continuous generation. Try it at lucy.decart.ai.
- Decart released Lucy 2.5 for realtime video editing: while the feed is still playing, you can add, swap, or remove content, reshape scenes, and run effects.
- Version 2.5 brings two upgrades: more realistic frames and tighter prompt following.
- On the product page: sub-40ms latency (under 40 milliseconds), 100fps continuous generation, and 4x realtime inference.
- Use cases include ecommerce virtual try-on and livestream shopping, interactive streams, ad variants, and live product placement.
- Try it: lucy.decart.ai; product page: decart.ai/lucy
Decart just shipped Lucy 2.5
On July 17, 2026, AI company Decart announced Lucy 2.5 on X. The company calls it its strongest realtime model yet.
In one line: Lucy can rewrite the picture while the video is still playing and still being generated. You don’t wait for a full render, then open a NLE and cut frame by frame.
What you can actually change on screen
Four core moves: add, swap, remove, and run effects.
Higher realism and visual fidelity; better prompt adherence (when you ask for a change, it tracks closer).
Classic path: record → import → fix frame by frame → export and ship. Lucy aims for the opposite: while the stream is still live, a prompt or instruction already rewrites the current frame—more like swapping props on set mid-show than sending the tape back to the edit bay.
What counts as “realtime”
Inference runs on Decart’s own DOS stack. The numbers below come from the product page, not our own benchmarks.
Elsewhere in the FAQ you also see figures like “within 200 milliseconds, 22FPS (512×768).” Those sit in a different section from the headline sub-40ms / 100FPS claims. The public pages don’t spell out how the two measurement setups differ. Keep this in mind: the lead marketing numbers are sub-40ms and 100FPS; end-to-end feel is whatever your own trial shows.
Lucy runs on Decart’s DOS. The company describes it as in-house inference and training infrastructure built for realtime workloads—the stack behind those latency and frame-rate claims.
Where it fits
Three main lanes: ecommerce, livestreaming, and ads. There are also social and gaming demos.
LOOOK.AI founder Dmytro Kornilov, roughly: Decart’s SDK can deliver realtime consumer experiences on the retail floor without tanking quality. That’s partner endorsement, not an independent lab report.
How to try it
The company calls Lucy the only realtime video model and points to a try link. That’s their own positioning line.
Official posts also say pricing will keep moving lower, but this launch doesn’t lock in package prices. Treat the try page and sales conversations as the source of truth on cost.
Meet Lucy 2.5, our most advanced realtime model yet. Lucy edits video in real time, now with more power and control. Decart official X, 2026-07-17
Primary sources: Decart launch thread (2026-07-17); product page decart.ai/lucy; try link lucy.decart.ai. Capabilities, latency, frame rates, and industry claims all come from vendor materials. The FAQ also mentions 200ms / 22FPS, without clarifying whether that setup matches the headline sub-40ms / 100FPS numbers.