LTX 2.3 MSR Multiple Subject Reference Video Generation

https://www.runninghub.cn/model/public/2064929870030921729

Multiple Subject Reference
LTX 2.3 Multiple Subject Reference

This model, based on Multiple Subject Reference (MSR), proposes a novel multi-reference image video generation scheme. This approach does not require additional encoder branches or feature fusion modules but instead converts multiple static reference images into pseudo-video sequences, placing them in the same feature representation space as the target video.

Instructions
This LoRA model needs to be used with the ComfyUI Licon MSR plugin for ComfyUI. The model file includes example workflows for quick testing and debugging.

Core Features
Multi-Reference Visual Memory
Token-level reference information retention: Encodes multiple reference images into video latent space features, preserving fine visual details at the token level rather than compressing them into a single feature embedding.
Native self-attention retrieval: Target video tokens can directly retrieve reference image tokens through the model's native self-attention mechanism without modifying the model architecture.
Contextual condition constraints: Reference content is integrated into the main token sequence as "visual memory" rather than as independent external conditional input.

Flexible Reference Image Combination Capability
Supports input of 2–5 reference images, adaptable to generation scenarios of varying complexity.
Each reference image can carry differentiated semantic information:
Subject identity characteristics
Props / object details
Scene and background
Local textures
Multi-angle views
Functional capabilities

Cross-Reference Identity Retention
Can simultaneously retain the subject features of multiple reference images in the generated video:
Merge multiple characters from multiple reference images
Combine characters and objects for generation
Blend objects and scenes for creative output

Content Combination Based on Associative Relationships
In addition to basic identity retention, the model can reorganize reference content based on associative logic described in text:
Action interactions (e.g., handing over, picking up, shoving)
Spatial positional relationships (e.g., left-right distribution, foreground/background layers)
Temporal event logic (e.g., start → process → result)

Selective Extraction of Cross-Reference Attributes
The model can intelligently select corresponding visual attributes from different reference images for fusion:
Facial features from reference image A, clothing features from reference image B
Object subject from one reference image, pose/position from another reference image
Background elements from scene-type reference images

Usage Tips (V1 Version)
Prompt crafting: Descriptions of reference images should be concise and precise. Over-description or insufficient information may lead to reduced image consistency.
High dynamic frames: It is recommended to set the frame rate to 50 frames/second to ensure smooth and coherent dynamic frames.
Generation stability: Typically, repeating sampling 2–3 rounds can yield accurate final results.