One workflow for multi-image reference video generation on LTX 2.3, with a three-stage render enhancement pass.
What it does
You supply a set of reference images and a prompt, and the graph produces a video where the subjects
come from the references. MSR (multi-subject reference) keeps them separate instead of blending them
into an average face.
Three-stage rendering is applied on top: structure pass, refinement pass, texture pass. Each stage
gets a progressively smaller denoise budget so the later stages adjust rather than rebuild.
Why multi-stage beats one big pass here
Reference-conditioned video accumulates drift. A single long pass tends to lose the references
halfway through. Splitting the render lets you re-anchor on the references between stages, which is
what keeps the subject looking like the subject at the end of the clip.
Running it
ComfyUI export, LTX 2.3 models required. The cloud instance on RunningHub already has them.
Running Hub links
Files
Workflow JSONs and covers: https://huggingface.co/AI-KSK/ksk-video-262-comfyui-workflows
Video walkthrough (Chinese): https://youtu.be/P-D1niNdJ4w
Bilibili: https://www.bilibili.com/video/BV1xsJw6YEg6/
Notes
- ComfyUI exports that reference models present in the cloud instance; not self-contained packs.
- Tutorial material, not a product.
- Independent AI-KSK release. Not affiliated with or endorsed by the model authors or RunningHub.
Top comments (0)