Decart's Lucy 2.5 Edits Live Video in Real Time — No Render Queue Required
Decart's Lucy 2.5 edits live video in real time at 30fps with near-zero latency, no render queue required. Here's how it actually works, and who it's for.
Decart released Lucy 2.5 on July 16, a real-time AI model that edits live video as it streams, at 30 frames per second, 1080p resolution, and near-zero latency. There's no render queue, no post-production pass. The edit happens to the frame as it's being captured and displayed.
Swap a shirt, change a character's face, restyle an entire environment into something else, all mid-stream, with the change visible in the same second it's requested. Decart's own framing calls it "Live AI," and it's a genuinely different category of tool than the generation-focused platforms that have dominated AI video coverage this year.
How It's Actually Different From Runway, Kling, or Veo
Every AI video model BRC has covered recently, Runway's Gen-4.5, Kling 3.0, Veo 3.1, Seedance 2.5, generates video from a prompt or reference material, then hands you a finished clip. Lucy 2.5 does something categorically different: it modifies video that's already playing, live, frame by frame, as a continuous stream rather than a discrete output.
That distinction matters more than it might sound. A generation model is a production tool used before or after the fact. Lucy 2.5 is closer to a live instrument, something you direct in real time while content is actively being created, streamed, or consumed.
The Technical Problem It Actually Solves
The hardest part of live AI video editing has always been temporal drift: an edited character's eye color shifts, clothing textures blur, facial structure subtly changes as a stream goes on, because each frame's edit compounds error from the frame before it.
Decart's answer is something the company calls Self-Anchoring. A few seconds into a stream, the model captures a snapshot of its own generated output and uses that as its new reference point, rather than continuing to build off the original unedited frame. Because the anchor comes from the model's own output distribution rather than a static source image, the guidance signal stays sharp instead of gradually drifting. In practice, Decart says edits remain stable for minutes at a time, even when a subject leaves frame and re-enters.
What It Can Actually Do
Lucy 2.5 supports character and face swaps, clothing and product changes, full environment restyling, added visual effects, and clean object removal, where the model reconstructs the newly exposed background rather than leaving a visible gap. Every edit type accepts a text prompt, a reference image, or both together.
The company has an SDK for JavaScript, Python, and Android, plus availability through fal.ai's serverless platform, meaning creators don't need to manage their own GPU infrastructure to integrate it into a live production.
Who This Is Actually For
The realistic use cases Decart is targeting are commerce (live virtual try-ons and product placement inserted into a stream in real time), advertising (one piece of live content becoming many localized or personalized variations without separate renders), and streaming itself (audiences influencing or transforming a creator's environment mid-broadcast).
For BRC's audience specifically, the more experimental application is real-time visual style transformation during a live shoot or stream, restyling an environment, applying VFX, or testing looks without waiting on a render pipeline. It's not a replacement for high-fidelity post-production work; the visual quality trade-off for real-time speed is real. It's a genuinely new tool for live and interactive content that didn't have a good AI-native option before.
Competitive Context
So @DecartAI just released Lucy 2.5, a real-time AI video editing model that lets you swap outfits, add VFX, relight scenes, and transform into different characters live.
— Curious Refuge (@CuriousRefuge) July 30, 2026
The visuals still have room to improve, but the real-time performance makes this one of the most exciting AI… pic.twitter.com/AoB1yOZbbM
Decart isn't a first-time player here; the company also built Oasis, an interactive real-time world model, and has raised $300 million from Radical Ventures, Sequoia, and NVIDIA, positioning itself specifically as infrastructure for real-time AI rather than a generation-focused competitor to Runway or Kling. That distinction is likely why direct competitors in this exact category are thin right now; most of the AI video field has concentrated on generation quality and length rather than live, in-stream editing.
The Signal in the Noise
The visuals still have real room to improve, and that's worth being upfront about; this isn't a replacement for a carefully composed, high-fidelity generation from Veo or Runway.
What makes Lucy 2.5 worth paying attention to isn't output polish, it's that the category itself, editing video while it's live rather than after it's rendered, barely existed as a usable product before this. Pricing is genuinely accessible for testing, at roughly $0.02 per second through the API, low enough to experiment with before committing to a production use case.
