DEV Community

AI-KSK
AI-KSK

Posted on

LTX-2.3 MSR Multi-Image Reference: IC-LoRA, Visual Memory and Three-Stage Consistency in One Workflow

#comfyui #ai #tutorial

Multi-image reference for video is the turning point this episode is about: LTX-2.3 MSR takes several references, keeps them consistent over a long render, and does it through an IC-LoRA-based route with three-stage rendering.

Evidence note: no workflow JSON was shipped and parameters were not parsed; the route and its features come from the episode description and the verified workflow page.

What "real" multi-image reference means

Many graphs accept several images but only use the first one meaningfully. MSR is presented as genuine multi-reference: the references participate in the generation, and visual memory carries their details across stages so identity does not drift when the render gets long.

The components

  • MSR (multi-image reference) — several references, not one image plus a prompt.
  • IC-LoRA universal route — the in-context LoRA path that makes reference conditioning work in a general graph rather than a bespoke one.
  • Visual memory + consistency constraints — the mechanism that keeps a character/prop stable across the sequence.
  • Three-stage rendering — structure, refinement, finish; each stage can reuse the same references.

When to use it

  • A character appears across several shots and must not drift.
  • A prop or costume has to stay identical between shots.
  • You need a long clip from references rather than a single still.

How to run

  1. Open the workflow: https://www.runninghub.ai/post/2065743626793209857/?inviteCode=rh-v1111
  2. Load the reference images.
  3. Write the shot prompt; set duration.
  4. Run the three-stage chain; inspect the mid-stage before the finish pass.

Limits

  • Multi-reference costs memory and time; more references is not automatically better.
  • Consistency depends on reference quality — conflicting references confuse the memory.
  • No JSON/parameters parsed for this episode.
  • Community workflow; LTX licensing follows upstream.

Episode resources

Top comments (0)