LTX-2.3 MSR V2: Multi-Subject Reference LoRA Gets Major Quality Upgrade
LiconStudio releases V2 of the popular Multiple Subject Reference LoRA for LTX-2.3, with improved consistency, stability, and scene logic based on community feedback.
LiconStudio has released V2 of their Multiple Subject Reference (MSR) LoRA for LTX-2.3, bringing significant improvements across consistency, stability, and scene logic based on extensive community feedback from V1.
While V1 was already a popular solution for multi-reference video generation — keeping track of up to 5 reference subjects in a single video — V2 addresses the most common pain points reported by users: identity drift, visual artifacts in complex compositions, and unnatural subject interactions.
What's New in V2
The update targets three core areas:
1. Improved Consistency
- Better preservation of character identity, clothing, objects, and scene details
- More consistent appearance across frames
- Improved alignment between multiple reference images and the generated video
- Reduced identity drift and reference attribute loss
2. Improved Stability
- More reliable results across repeated sampling runs
- Reduced visual artifacts, flickering, and temporal inconsistencies
- More stable generation in complex multi-subject compositions
- Improved handling of motion and interactions between subjects
3. Improved Scene Logic
- Better understanding of spatial and action relationships described in prompts
- More natural subject positioning and interaction
- Improved temporal progression from the beginning to the end of a video
- More coherent composition of characters, objects, and backgrounds
How MSR Works
Unlike traditional multi-reference approaches that require additional encoder branches or fusion modules, MSR transforms multiple static reference images into a pseudo-video sequence. This sequence shares the same representation space as the target video, allowing reference tokens to be accessed through the model's existing self-attention mechanism — no additional architectural components needed.
The LoRA supports 2 to 5 reference images, each providing complementary semantic information such as subject identity, object details, scene background, or multiple viewpoints.
Usage
The LoRA requires the ComfyUI-Licon-MSR custom node plugin, which handles the creation of fixed-frame MP4 reference videos from multiple subject images. A sample workflow is included in the model files for easy experimentation.
Tips for Best Results
- Use concise but accurate descriptions of each reference image — both excessive and insufficient descriptions can reduce consistency
- Clearly describe the role of each referenced subject, object, or scene in the target video
- For high-motion scenes, 50 fps is recommended for smoother motion coherence
- Complex multi-subject interactions may still benefit from multiple sampling runs
Availability
The V2 LoRA is available on Hugging Face under the Apache 2.0 license:
Model: LiconStudio/LTX-2.3-Multiple-Subject-Reference
Plugin: ComfyUI-Licon-MSR
Base model: Lightricks/LTX-2.3