Lucy 2.5 Live AI Redefines Real-Time Video

Decart AI has launched Lucy 2.5, a significant upgrade to its live AI video editing platform. This new iteration pushes real-time video manipulation to 30 frames per second at 1080p resolution, boasting near-zero latency. The system continuously generates each frame, reacting dynamically to scene changes without relying on pre-rendered footage. This advancement allows for unprecedented real-time control over video content, as detailed by Decart AI.

Transforming Video in Real-Time

Lucy 2.5 enables users to perform complex edits live, including swapping characters, altering clothing, adding physically accurate visual effects, modifying entire environments, and removing objects. These transformations are driven by text prompts and reference images, all processed instantaneously. New features enhance edit control, temporal stability, and generalization, making the results more realistic and consistent.

The system employs a dedicated model for propagating image edits across video frames and utilizes prompts aligned with generic scene descriptions for emergent generalization. A coarse-to-fine prompting approach offers users both high-level instructions and granular control over details like position, style, color, and size.

Physically Aware Visual Effects and Global Scene Control

Visual effects like slime, sand, and water now behave with physical realism. The platform also supports global style changes, allowing entire scenes to be re-themed while preserving structural integrity and subject identity. This means a sunny street can be transformed into a cyberpunk dystopia, maintaining the original scene's geometry and subjects.

Subject replacement is seamless, with the model understanding original volumetric space to ensure new elements integrate correctly with shadows and physical interactions. Object removal is equally refined, with the AI reconstructing the background consistently over time.

Persistent Edits with Self-Anchoring

A key innovation is Self-Anchoring, where Lucy 2.5 adopts its own generated output as a new reference point. This ensures edits remain locked in, even minutes into a stream, overcoming common error accumulation in autoregressive models. Identity, geometry, and texture are maintained over extended periods, even as subjects move or leave the frame.

This persistence is further enhanced by Smart History Augmentation, ensuring stability across dramatically longer video horizons.

Optimized for Speed

Achieving this level of real-time performance required significant optimization of the inference stack. Decart AI implemented low-precision quantization (MXFP8 and NVFP4) for up to 4x speedups. Dynamic sparse attention algorithms and deeper kernel fusion further reduce latency by optimizing computational efficiency and minimizing hardware transaction overhead.

These technical advancements ensure that model quality is maintained while pushing performance boundaries.

Applications for Live AI

Lucy 2.5 unlocks new possibilities for interactive experiences across various industries. In commerce, it enables live virtual try-ons and personalized product showcases. Advertising can leverage real-time localization and creative adaptation for dynamic ad variations.

Streamers can offer new ways to engage audiences by allowing real-time influence on the broadcast. Gaming can blend interactive gameplay with the physical world, creating immersive experiences. Lucy 2.5 positions Live AI as a critical business advantage for companies building the future of digital interaction.