LTX25 IC-LoRA Loader (MSR)

MSR - multiple-subject reference - is the LTX2.5 family's trick for keeping several characters consistent at once: one IC-LoRA checkpoint holds both the regular weights and a set of learned reference-slot embeddings, and you condition generation on a few reference images that each slot into one position. Normally you'd need the separate ComfyUI-LTX2.5-MSR package just to load one. This node does the same job with no extra install, because it ships inside the same pack as the LTX25 Reference Params node it feeds.

What it does

Two things at once, which is why it's a loader and not a "params" node:

1. It applies the LoRA's regular weights to the diffusion model, exactly like any other LoRA loader.
2. It extracts the learned reference-slot embedding - a Fourier-MLP that maps a slot index to an embedding - and hands it out the second output so LTX25 Reference Params can embed each reference still into its own slot.

Its msr_parameters output uses the same wire type (LTX_MSR_REFERENCE_PARAMETERS) as the ComfyUI-LTX2.5-MSR package's IC-LoRA Loader, so you can swap one for the other without rewiring anything downstream.

The inputs and outputs

Three required inputs, all straightforward:

• model - the LTX2.5 diffusion model the LoRA is applied to.
• lora_name - the MSR multi-reference LoRA checkpoint, picked from your loras folder.
• strength_model (default 1.0) - how strongly the regular LoRA weights affect the model. LoRA strength discipline is a genuine thing with LTX; an overweighted LoRA is the usual suspect when faces start looking "diseased" or identity melts, so tune this before you blame the rest of the pipeline.

Outputs: model (the model with the regular weights applied - feed this into the upscale/sampling stage) and msr_parameters (wire it into LTX25 Reference Params). Set strength_model to 0 and the regular weights are skipped but the slot embedding is still extracted, if you ever want references without the LoRA's style.

Installing

Part of bbaudio-2025/Comfyui-MMH3-UltimateUpscale. Via ComfyUI Manager (search "Comfyui-MMH3-UltimateUpscale") or:

cd ComfyUI/custom_nodes
git clone https://github.com/bbaudio-2025/Comfyui-MMH3-UltimateUpscale

Restart ComfyUI. There are no pip dependencies - the loader is self-contained, with the slot-extraction logic copied into the pack so you never need the ComfyUI-LTX2.5-MSR package installed. You do need a recent ComfyUI and the MSR LoRA checkpoint in loras.

Troubleshooting

• "This LoRA does not contain reference_slot_embedding weights" - you've pointed it at a regular (non-MSR) LoRA. This loader validates, and it's picky on purpose: a plain LoRA has no slot embeddings to extract, so it fails loudly rather than silently doing half a job.
• "reference_slot_embedding_ tensors are missing"* - the checkpoint declares slot embeddings in its metadata but the actual tensors aren't there; the file is likely truncated or mis-exported. Re-download it.
• "reference_downscale_factor != 1 … not supported" - the chunked pipeline in LTX25 Ultimate Upscale only works with factor-1 MSR checkpoints. The loader reads the factor from the LoRA's metadata and surfaces it; use a factor-1 checkpoint.
• References don't stick - check ref_strength in the Reference Params node and the reference-image order (batch order = slot order, and slot 1 starts the numbering).

One last thing: the pack's LTX25 line is flagged "highly experimental" by the author, and MSR itself is a young, fast-moving part of the LTX ecosystem. When a checkpoint misbehaves, it may genuinely be the LoRA - check its own release notes before assuming the loader is at fault.