Descriptions:
This tutorial covers LTX Video 2.3 MSR — Multiple Subject Reference — a video generation model that maintains visual consistency across multiple reference subjects simultaneously inside ComfyUI. Unlike single-image reference approaches, MSR packages multiple reference images into a short “reference video” using a dedicated node called LiconMSR, creating a visual memory the model consults during generation. The model itself is implemented as an IC LoRA, downloadable from Hugging Face and dropped into a standard LTX 2.3 installation.
The guide walks through the full ComfyUI setup: installing the official LTXVideo node package, the ComfyUI-Licon-MSR extension, and configuring the dual-chain workflow that combines LTXV Add Guide Multi — which provides a strong visual starting point — with LTX Add Video IC LoRA Guide, which converts the MSR reference video into positive and negative conditioning for the sampler. A key technical detail is the frame_count rule inside LiconMSR (values must be multiples of four plus one), which is separate from the output video’s total frame count — the tutorial uses 241 frames for roughly a ten-second output.
Five test scenarios benchmark the model’s capabilities: two-character consistency to verify faces don’t contaminate each other, character sheet references (three-view inputs) for improved dimensionality, character-plus-prop setups using a woman-and-luxury-car combination, storyboard-based scene control for precise motion direction, and a maximum-stress test using all five reference slots simultaneously. Results show strong subject-identity preservation across most scenarios, with expected limitations in fine details like hand rendering.
📺 Source: Veteran AI · Published June 11, 2026
🏷️ Format: Tutorial Demo
