Lucy 2.5: Raising the Bar for Live AI

‍Lucy 2.5 pushes real-time AI editing further than ever before. More than a higher-quality video generation model, Lucy is a live system capable of transforming the visual world at 30 FPS and 1080p resolution, with near-zero latency. Every frame is generated in real time, responding continuously to motion, lighting, and the evolving scene, never relying on pre-rendered video.

With Lucy 2.5, users can seamlessly swap characters, change clothing, add physically-aware visual effects, transform entire environments, remove objects, and precisely edit scenes using text prompts and reference images, all while the video is still streaming live. New advances in edit control, temporal stability, and generalization make these transformations more realistic, more consistent, and more reliable than ever before.

— The Decart Team

More Natural, More Precise, More Powerful Edits

• General
  
  • We used a dedicated model for propagating image edits to video edits
  • We used prompts that are aligned with generic scene descriptions to increase emergent generalization
• Fine control
  
  • We used a coarse-to-fine prompting approach
    
    • Coarse - simple edit instruction
    • Fine - details about position, style, color, size, etc.
• VFX (visual effects) are improved. Elements like slime, sand, water, fire, etc.
• Global style supports changing the entire scene while preserving structure and identity
• Remove functionality is supported

Lucy 2.5 substantially improves generalization for general editing. By aligning our training prompts with universal scene descriptions, we bypass rigid instruction sets in favor of deep, emergent generalization. We paired this with a coarse-to-fine caption control hierarchy, allowing users to guide the model with simple, high-level edit instructions or drill down into precise, fine-grained details regarding position, style, color, and size.

To achieve this level of emergent capability, a model requires training data at a scale and complexity that does not organically exist. To bridge this gap, we engineered a massive synthetic corpus of video editing data. Starting with a vast foundation of high-quality image-to-image transformations, we used specialized models trained internally to propagate static image edits across full video sequences. This pipeline effectively translates single-frame modifications into rich, temporally consistent video editing pairs used to train our model.

As a result:

• VFX behaves physically: Volatile and complex elements like flowing water, fire, shifting sand, and viscous slime interact naturally within the environment.
• Worlds transform globally and scenes can be re-themed: The aesthetic of an entire video can be radically reimagined. Users can transition the fundamental setting of a scene, such as shifting a sunny city street into a neon cyberpunk dystopia, while strictly preserving the underlying structure, geometry, and identity of the original subjects.
• Subjects can be seamlessly replaced: When swapping one object for another, the model understands the original volumetric space, ensuring the new element matches existing contact points, shadows, and physical interactions.
• Objects can be cleanly removed: When an element is erased, the model naturally reconstructs the newly exposed background geometry, and does so consistently over time.

‍

Edits That Stay Locked In

Lucy 2.5 introduces Self-Anchoring. Shortly into a stream, the model adopts a snapshot of its own generated output as its new reference, one that already reflects the edit as realized in the actual scene, under the actual lighting, framing, and motion. Because the anchor comes from the model's own distribution, the guidance signal stays sharp and informative for the entire rollout instead of slowly diverging due to the common error accumulation present in autoregressive models. Importantly, since we also train the model under the same self-anchored conditions it faces at inference, it learns not merely to tolerate its own output as a reference; it learns to exploit it.
As a result:

• Edits persist. A character swap or costume change made in the first seconds of a stream is still locked in minutes later.
• Identity survives motion. Subjects can leave frame, turn around, and return without the edit dissolving or morphing.
• Stability compounds instead of decaying. Combined with Smart History Augmentation, streams hold identity, geometry, and texture over dramatically longer horizons.

Faster Than Ever: Behind the Inference Stack

Real-time performance is a moving target: every gain in model capability raises the bar for the stack beneath it. To push latency down even further, we improved on our previous inference stack in three key ways:

• Low-precision quantization. We employ multiple custom quantization techniques to run inference in MXFP8 and NVFP4, unlocking up to 4x speedups on compute-bound operations thanks to native hardware support.
• Dynamic sparse attention. Attention dominates our Transformer's forward pass, so we developed a custom sparse attention algorithm backed by tailored kernels. It dynamically identifies which computations in each attention call are redundant and prunes them with no loss in model accuracy.
• Deeper kernel fusion. More aggressive fusions and improved megakernels let us better utilize the whole chip, further reducing the launch overhead and costly HBM transactions that are critical in low-latency inference.

These optimizations build on our existing foundations, letting us keep pushing performance while leaving model quality untouched.

Real-World Applications

Lucy 2.5 unlocks a new generation of real-time interactive experiences. Because every frame remains editable while it is being generated, live video becomes a dynamic medium that continuously adapts to users, audiences, and context.

With Lucy 2.5:

• Commerce: Shopping becomes interactive with live virtual try-ons and personalized product experiences.
• Advertising: One ad becomes infinitely many variations through real-time localization, personalization, and creative adaptation.
• Streaming: Streamers unlock new ways to engage their communities, allowing audiences to influence the experience in real time.
• Gaming: Games break beyond the screen, blending interactive gameplay seamlessly with the physical world.

Whether you’re building the future of commerce, advertising, entertainment, gaming, or media, Lucy 2.5 turns Live AI into your next business advantage.

Explore Lucy

Experience Live AI