Summary
Veteran AI walks through MSR V2 (Multiple Subject Reference version 2), a ComfyUI-based technique for generating AI videos that maintain consistent character identity across shot cuts and complex multi-subject scenes. The system is built around LTX 2.3 and the ComfyUI-Licon-MSR extension, accepting up to four reference images plus a required background image, and organizing them into a reference sequence—65 frames by default—that guides final video generation at up to 241 frames (approximately 10 seconds).
The tutorial covers three core workflow nodes: reference image slots (each assigned a single clear responsibility—character, prop, background), the LiconMSR node that converts those images into a reference sequence, and PromptRelayEncode, a model-specific prompt encoder that asks users to first describe each reference image’s role and then write the action sequence in initial-state, process, and final-state order. This structured prompting makes debugging easier by isolating which stage produces unexpected results. The video also covers the RunningHub platform as an accessible online alternative for running these ComfyUI workflows without local hardware setup.
Capacity tests show clean results for one to three characters, with some dropout risk at four. Advanced demonstrations include attribute assembly—combining a face reference, a costume, and a prop from three separate images into one coherent character—narrative event generation (a seed sprouting and blooming), and an object handoff sequence tracing a chip from a tray to tweezers to a robotic hand, illustrating both the technique’s strengths and its current limits around fine-grained prop motion fidelity.
📺 Source: Veteran AI · Published July 17, 2026
🏷️ Format: Tutorial Demo
