LTX-2.5-Multiple-Subject-Reference

<h1 align="center">Licon MSR V1 for LTX-2.5</h1> <p align="center"> <a href="https://github.com/liconstudio/ComfyUI-LTX2.5-MSR"> <img src="https://img.shields.io/badge/GitHub-Code-181717?style=for-the-badge&logo=github&logoColor=white" alt="GitHub Code"> </a> <a href="https://arxiv.org/abs/2609.18393"> <img src="https://img.shields.io/badge/arXiv-Paper-B31B1B?style=for-the-badge&logo=arxiv&logoColor=white" alt="arXiv Paper"> </a> </p> <p align="center"> <a href="https://arxiv.org/abs/2609.18393"> <b>MSR: Multiple Subject Reference for Video Generation</b> </a> </p>

What's New in V2

V2 improves visual consistency, character motion, and camera behavior, and adds audio reference support.

• Reduced clipping and interpenetration.
• More varied camera movement and framing.
• More natural, less rigid character motion.
• More coherent shot transitions, with fewer erratic cuts.
• Reduced unintended duplication of reference objects.
• Audio references for up to two speakers, paired with their corresponding visual references.

Current Issue

Environmental sound generation currently requires background music in the voice-reference audio. Without it, environmental sounds may not be generated correctly.

Overview

Licon MSR V2 is a multi-reference LoRA trained for LTX-2.5, with support for both visual and audio references.

It uses the Multiple Subject Reference (MSR) approach to encode multiple reference images as visual tokens in the same latent space as the target video. Each reference is assigned a learned slot embedding and a distinct negative temporal position, allowing target video tokens to retrieve character, clothing, object, and scene information through the model's native self-attention layers.

Key Features

• Supports up to five reference images
• Supports audio references for up to two speakers
• Preserves multiple characters, clothing, objects, and backgrounds
• Learned slot embeddings distinguish different references
• Native self-attention retrieval of reference details
• Supports multi-subject and subject-object composition
• Designed specifically for the LTX-2.5 architecture

Usage

ComfyUI inference requires ComfyUI-LTX2.5-MSR. A sample workflow is included in the plugin repository.

For audio references, encode each clip with the native LTXV Audio VAE Encode node and connect its LATENT output to audio_ref1 or audio_ref2 on the main Multi-Reference Guide. These inputs correspond to pic1 and pic2, respectively. Leave an audio input disconnected when that subject has no audio reference.

Usage Tips

• Describe each reference image clearly in the prompt.
• Use consistent labels such as Image 1, Image 2, and Image 3.
• Clearly specify subject actions and spatial relationships.
• Specify which reference provides the character, object, clothing, or background.
• Match each audio reference to the corresponding subject image.

V2 Example

<table> <tr> <th align="center" colspan="3">Example 04</th> </tr> <tr> <td align="center" width="220"> <b>Reference Image 1</b><br><br> <img src="https://huggingface.co/LiconStudio/LTX-2.5-Multiple-Subject-Reference/resolve/main/Validation-V2/04/01.png" width="220" alt="Reference Image 1"> </td> <td align="center" width="220"> <b>Reference Image 2</b><br><br> <img src="https://huggingface.co/LiconStudio/LTX-2.5-Multiple-Subject-Reference/resolve/main/Validation-V2/04/02.png" width="220" alt="Reference Image 2"> </td> <td align="center" width="220"> <b>Reference Image 3</b><br><br> <img src="https://huggingface.co/LiconStudio/LTX-2.5-Multiple-Subject-Reference/resolve/main/Validation-V2/04/03.png" width="220" alt="Reference Image 3"> </td> </tr> <tr> <td align="center" colspan="3"> <video src="https://huggingface.co/LiconStudio/LTX-2.5-Multiple-Subject-Reference/resolve/main/Validation-V2/04/V2.mp4" controls width="660"></video> <br><b>Licon MSR V2</b> </td> </tr> </table>