Home Creators Posts Import Register Favorites Logout

Downloads

Content

Video : https://youtu.be/EFm4z0ZF20M

July 2026

Hey everyone. New video is up, and this one is about the MSR V2 (Multi-Subject Reference) LoRA update for LTX 2.3.

The V2 model from Licon Studio improves character consistency, object stability, and scene logic compared to V1. You feed it one reference image — a character sheet, a screenshot, anything — and it generates multiple scenes while keeping the outfit, accessories, and body type consistent across different camera angles and shots.

I tested it with a music video demo. The character has tattoos on both sides of the neck and hands. Even in far shots, close-ups, and camera orbits, the tattoos stay recognizable. The outfit — a leather jacket with neon light straps — stays consistent across every angle. The colorations are noticeably better than V1.

But here is the honest part: facial recognition is still the limitation. Even with a close-up reference image, the generated face does not match the reference exactly. This is not a V2 problem — it is a reference-to-video problem across all AI video models, including frontier paid ones. For exact character face identity, you need to train a character LoRA with Krea 2 or similar tools.

The free workflow uses LTX 2.3 22B + MSR V2 LoRA + LiconMSR node for character sheet input.

Patreon Supporters , please check out the member post for additional contents.

— Benji

Resources:

LiconStudio/LTX-2.3-Multiple-Subject-Reference

https://huggingface.co/LiconStudio/LTX-2.3-Multiple-Subject-Reference

Files

(No title)

Comments

No comments found for this post.