Licon Studio released a LoRA adapter for working with multiple objects in LTX-2.5

The Licon Studio team introduced LTX-2.5-Multiple-Subject-Reference — a specialized LoRA adapter designed to expand the capabilities of the base LTX-2.5 video generation model. The tool allows the system to maintain visual identity for up to five different objects, characters, or background elements simultaneously.

The technology works by encoding reference images as visual tokens. During generation, the model uses a self-attention mechanism to match these tokens with the target video sequence. This provides high accuracy in transferring object characteristics from the source files to the final animation without losing detail.

The solution is targeted at users who require strict control over character consistency in the video stream. The adapter is optimized to work with the LTX-2.5 architecture, allowing it to be integrated into existing generation pipelines without the need to retrain the main model. The source code and model weights are available for download on the Hugging Face platform.

The emergence of such tools simplifies the creation of videos with fixed characters, which previously required complex inversion methods or frame-by-frame processing. Using a multi-reference approach reduces the time spent preparing prompts and increases the predictability of the result when working with multiple subjects in a single frame.