Licon MSR for LTX-2.5

Licon MSR for LTX-2.5

Keeps characters, clothing, objects and backgrounds consistent in generated video by feeding the model up to five reference images. A LoRA for LTX-2.5 in ComfyUI, with optional audio references for two speakers.

Made by LiconStudio on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Licon MSR is a LoRA trained for the LTX-2.5 video model. It lets you hand the model several reference images, so a generated clip keeps the same characters, clothing, objects and scenes. The newer V2 release also accepts audio references for up to two speakers.

Who it suits: Video creators and ComfyUI users who want the same subjects to appear across generated shots. It fits anyone who already runs LTX-2.5 and wants more control over who and what shows up.

Industry | Who uses it | How they use it | Example
Video production | Short-film and clip makers | Supply character and scene references to keep a cast consistent | A clip with two characters and a set built from three images
Marketing | Content creators | Place a product and a person into one generated shot | A product shown with a presenter reference
Game and animation | Concept artists | Test how characters and props move in a scene | Previewing a character in several camera angles
Research | Video generation researchers | Study multi-reference conditioning in a video model | Comparing results with different reference labels

You can skip it if you do not use LTX-2.5 or ComfyUI, since it is built for that model. Use only faces and voices you have the right to use.

Where to use it:

• ComfyUI, with the ComfyUI-LTX2.5-MSR plugin installed
• Windows, macOS or Linux with Python, to fetch the files
• Workflows that already run the LTX-2.5 base model

Use case | Who | What they get back
Multi-character video | Video creators | A clip that keeps each referenced character
Subject and object composition | Designers | A scene combining people, objects and a background
Talking subjects | Creators | Speech paired with the matching subject image

How to set it up:

• Python, to install the Hugging Face client
• ComfyUI and the LTX-2.5 base model
• The ComfyUI-LTX2.5-MSR plugin, which the card says ComfyUI needs

1. Install the Hugging Face client.
2. Download the model files to your computer.
3. Install the ComfyUI-LTX2.5-MSR plugin in ComfyUI.
4. Open the sample workflow that comes with the plugin.

pip install -U huggingface_hub
hf download LiconStudio/LTX-2.5-Multiple-Subject-Reference

How to use it: Load up to five reference images in the sample workflow and describe each one in your prompt. For speech, pass each audio clip through the LTXV Audio VAE Encode node. Wire its LATENT result into the audio_ref1 slot or the audio_ref2 slot of the Multi-Reference Guide. Those slots match pic1 and pic2. Leave an audio input empty when a subject has no audio.

Tool | What it does
Multi-Reference Guide | Main node that takes your reference pictures and audio
LTXV Audio VAE Encode | Encodes each audio clip for the guide node
audio_ref1 and audio_ref2 | Inputs that pair speaker audio with pic1 and pic2

Example: You load three reference images, label them Image 1, Image 2 and Image 3 in your prompt, and say what each one is, such as the character, the object and the background. You also describe the actions and where things stand. The model returns a video that keeps all three.

For newcomers and experts: If you are new, open the sample workflow and start with two reference images and a short, clear prompt. Name each image in the prompt and match any audio to the right subject.

If you know the field: the reference images are turned into visual tokens that share a latent space with the target video. Each gets a learned slot embedding and its own negative temporal position, so the video tokens fetch details through the native self-attention layers.

Reference images
Visual tokens with slot embeddings
Self-attention retrieval
Generated video

Good to know: V2 reduces clipping, adds more varied camera moves and gives more natural motion with fewer erratic cuts. One known issue is that environmental sound needs background music in the voice-reference audio.

• Limits: up to five reference images and audio for up to two speakers
• Known issue: without background music in the voice reference, environmental sound may not generate correctly
• License: Apache 2.0
• Cost to run: none named
• Data it reaches: model files downloaded to your computer, plus the reference images and audio you supply
• Not for: the card names none, so use only faces and voices you have the right to use

On Hugging Face, 14 November 2026

• Library: diffusers
• Downloads in the last 30 days: 14,885
• Likes: 110
• License: apache-2.0
• Based on: Lightricks/LTX-2.5
• Tags: diffusers, video-generation, multi-reference, ltx-2.5

Links

• Model: https://huggingface.co/LiconStudio/LTX-2.5-Multiple-Subject-Reference
• Files: https://huggingface.co/LiconStudio/LTX-2.5-Multiple-Subject-Reference/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download LiconStudio/LTX-2.5-Multiple-Subject-Reference

Then follow the usage in its model card. Nothing in this kit runs by itself.