LTX 2.3 MSR V2 vs V1: ComfyUI workflow and review

This review tests the MSR V2 LoRA for LTX Video 2.3 in ComfyUI, using a character reference image to generate multiple scenes. The presenter reports better outfit consistency, steadier objects and less flickering than V1. A music video demo shows clothing, hairstyles and approximate tattoo shapes across camera angles, though tattoo patterns do not match exactly.

The workflow accepts the V2 LoRA in place of V1. The presenter recommends higher frame counts in the LiconMSR custom node, such as 41, while noting that the demo used an older node version with a setting of 17. Improvements from updating the node and increasing settings remain an expectation rather than a demonstrated result.

For the comparison, the presenter keeps the text prompts, audio and other settings unchanged. V2 follows the reference character's appearance more closely in these examples, but still misses the exact face. A second test adds a close-up face reference and reaches a similar limitation. The presenter uses BF16 and estimates roughly 80 percent outfit and style consistency with high precision models; this is a personal assessment, not a measured benchmark.

For recurring characters that need a stable facial identity, the presenter recommends training a dedicated character LoRA or using face swapping. The review ends with the assembled music video.