Skip to main content
Back

Licon Studio's MSR V2 Lets LTX-2.5 Control Five Characters in one Shot

A slot-aware LoRA for Lightricks' LTX-2.5 lets you condition video generation on up to five reference images and two voice tracks.

  • Licon Studio released MSR V2, a multi-subject reference LoRA for Lightricks LTX-2.5.
  • Supports up to five reference images plus audio references for two distinct speakers.
  • Each reference gets a slot embedding and a negative temporal offset in shared latent space.
  • V2 reduces clipping, object duplication, and rigid motion versus the first version.
  • Runs through the ComfyUI-LTX2.5-MSR plugin with a sample workflow included.
  • Method detailed in the MSR paper; Apache 2.0 license, over 10,000 downloads.

Licon Studio has released a LoRA adapter that helps Lightricks’ LTX-2.5 video model preserve multiple characters, outfits, objects, and backgrounds within a generated shot. MSR V2 supports five visual references and two speaker-audio references through a slot-based conditioning system that uses LTX-2.5’s existing attention layers.

Component Details
Adapter Multi-reference LoRA for LTX-2.5
Visual inputs Up to five reference images
Audio inputs Up to two speaker references
Interface ComfyUI node pack
Adapter license Apache 2.0
Adoption More than 10,000 Hugging Face downloads at publication

The adapter’s Apache 2.0 license applies to the MSR V2 release. Deployments must also comply with the separate terms attached to the LTX-2.5 weights.

A separate slot for every source

MSR V2 encodes each image independently as a static clip, then represents it as a distinct group of latent tokens. Those tokens occupy the same latent space as the generated video, allowing the base transformer to process source images and target frames within one attention sequence.

A compact Fourier-feature multilayer perceptron converts each numeric slot identifier into an embedding. Slot-specific temporal offsets also shift the group’s rotary positional coordinates, giving every image a distinct negative temporal position and reducing confusion among visually similar sources.

During flow-matching training, clean source tokens are prepended to noisy target-video tokens. Flow matching is the objective that teaches the model to transform noise into a video sample. Training updates are calculated for the target tokens, while the source groups remain available as context.