Decart AI Updates Lucy 2.5 for Real-Time 1080p Video Editing
The updated video model introduces an anchoring trick to stop characters morphing between frames, backed by aggressive 4-bit quantization.
Startup Decart AI has released version 2.5 of its video-to-video editing model, Lucy. According to the company, the system modifies live 1080p video streams at 30 frames per second using text prompts or reference images without pre-rendering.
Real-time video generative models notoriously suffer from temporal flicker and "amnesia," where a subject walks out of frame and returns as a completely different person. To mitigate this, Decart AI introduced what it calls "Self-Anchoring." The model creates an initial reference representation of the scene at startup, using it as a baseline anchor so characters and objects maintain their visual identity across frames.
On the computational side, Decart AI claims a fourfold boost in generation speed compared to its previous iteration. The speedup relies on heavy low-precision quantization using MXFP8 and NVFP4 formats, combined with sparse attention and kernel fusion to cut memory bandwidth and kernel launch overhead.
Whether NVFP4 quantization degrades visual fidelity in complex, dynamic lighting remains unverified outside the company's demonstrations, as independent stress tests under live streaming conditions have yet to surface.
Source: habr
Primary source: Lucy
Comments
Help shape the next version: Add context or suggest a correction. AI review can add points toward a rewrite. Reviews and updates may take time; a full meter does not guarantee a new version.