Multiple Subject Reference LoRAs trained by SyFe on LTX-2.3 22B-dev. Multiple subject and scene images are encoded as reference-video latents so target tokens can retrieve them through native self-attention.

Checkpoints

Run	Data / construction	Rank	Steps	Status
`msr_plain_01`	Initial crop-fill references	128	5,000	Archived: crop-fill can crop full-body subjects
`msr_plain_02`	Correct white-canvas, never-crop subjects	128	5,000	Validated in the combined talking stack
`msr_plain_rebuilt01`	Rebuilt intermediate corpus	128	5,000	Superseded experiment
`msr_corpus36_run01`	35 shows, 31,500 true-bilingual samples	128	6,000	Recommended; deployed final

For msr_corpus36_run01, reference subjects must be contain-fit without cropping on a white canvas; scene references may be cover-fit. Prompts should begin with the ordered subject markers, for example [VISUAL]: char_1_person, char_2_person..., because marker order binds prompt subjects to reference slots.

The recommended checkpoint materially improves out-of-show costume, prop, hair, and scene adherence and removes memorized-cast substitution. Exact facial identity can still drift, especially when faces occupy few pixels. Two-subject reliability is not perfect.

The third-party Licon-MSR-V1 checkpoint is intentionally not included. Use is subject to the LTX-2 community license.

SyFeee/LTX-2.3-SyFe-MSR-LoRA

SyFe LTX-2.3 MSR LoRA Checkpoints

Checkpoints

SyFeee/
LTX-2.3-SyFe-MSR-LoRA