Instructions to use LiconStudio/LTX-2.3-Multiple-Subject-Reference with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use LiconStudio/LTX-2.3-Multiple-Subject-Reference with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("LiconStudio/LTX-2.3-Multiple-Subject-Reference", dtype=torch.bfloat16, device_map="cuda") prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k" image = pipe(prompt).images[0] - Notebooks
- Google Colab
- Kaggle
| license: apache-2.0 | |
| tags: | |
| - video-generation | |
| - multi-reference | |
| - LTX-2.3 | |
| base_model: | |
| - Lightricks/LTX-2.3 | |
| github: https://github.com/liconstudio/ComfyUI-Licon-MSR | |
| library_name: diffusers | |
| # Licon MSR V2 for LTX-2.3 | |
| ## What's New in V2 | |
| Compared with V1, **Licon MSR V2** introduces significant improvements in three key 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 | |
| ## Overview | |
| This model implements a novel approach to multi-reference video generation using **Multiple Subject Reference (MSR)**. Instead of introducing additional encoder branches or fusion modules, we transform multiple static reference images into a pseudo-video sequence that shares the same representation space as the target video. | |
| ## Usage | |
| This LoRA requires the **[ComfyUI-Licon-MSR](https://github.com/liconstudio/ComfyUI-Licon-MSR)** plugin for ComfyUI. A sample workflow is included in the model files for easy testing and experimentation. | |
| ## Key Features | |
| ### Multi-Reference Visual Memory | |
| - **Token-level reference preservation**: Multiple reference images are encoded as video latents, preserving fine-grained visual information at the token level instead of compressing them into a single embedding | |
| - **Native self-attention retrieval**: Target video tokens directly access reference tokens through the model's existing self-attention mechanism, with no additional architectural components required | |
| - **In-context conditioning**: References serve as visual memory within the main token sequence rather than as external conditioning inputs | |
| ### Flexible Reference Composition | |
| - **2 to 5 reference images**: Supports varying numbers of reference inputs with increasing composition complexity | |
| - **Complementary semantic roles**: Each reference image can provide different information: | |
| - Subject identity | |
| - Object or prop details | |
| - Scene or background | |
| - Local textures | |
| - Multiple viewpoints | |
| ## What It Can Do | |
| ### Identity Preservation Across References | |
| Generate videos in which multiple reference identities are simultaneously preserved: | |
| - Multiple characters from different reference images | |
| - Character and object combinations | |
| - Object and scene compositions | |
| ### Relation-Based Composition | |
| Beyond identity preservation, the model can compose references according to textual relationship descriptions: | |
| - Action interactions, such as handing, picking up, or pushing | |
| - Spatial relationships, such as left and right or foreground and background | |
| - Temporal event structures, such as start → process → result | |
| ### Cross-Reference Attribute Selection | |
| The model learns to selectively retrieve attributes from different references: | |
| - Face from reference A and clothing from reference B | |
| - Object identity from one reference and pose or position from another | |
| - Background elements from scene references | |
| ## Usage Tips | |
| - **Prompt description**: Use concise but accurate descriptions of the reference images. Both excessive and insufficient descriptions may reduce consistency. | |
| - **Reference roles**: Clearly describe the role of each referenced subject, object, or scene in the target video. | |
| - **High-motion scenes**: 50 fps is recommended for smoother motion coherence. | |
| - **Sampling**: Complex multi-subject interactions may still benefit from multiple sampling runs. | |
| ## V1 vs. V2 Comparison | |
| ### Comparison 1 | |
| | V1 | V2 | | |
| |:---:|:---:| | |
| | [▶ Play V1](Validition_V2/01/V1_1.mp4) | [▶ Play V2](Validition_V2/01/V2.mp4) | | |
| ### Comparison 2 | |
| | V1 | V2 | | |
| |:---:|:---:| | |
| | [▶ Play V1](Validition_V2/02/V1.mp4) | [▶ Play V2](Validition_V2/02/V2.mp4) |