Nodes/H3 Relay/H3 Relay · Interpolation Model Loader
ComfyUI Node

H3 Relay · Interpolation Model Loader

Where H3 Relay's frame interpolation model comes from

By akatz-ai·Created 28 days ago·Updated 8 days ago· 15
H3 Relay · Interpolation Model Loader
    • interpolation
    model_name
    manual_cache_revisionv1

    Every H3 Relay movie needs the finishing pass that turns 24fps LTX video into buttery 48fps+, and every such pass needs an interpolation model loaded first. H3RelayInterpolationModelLoader is that loader. It grabs a frame-interpolation checkpoint from ComfyUI's frame_interpolation folder and wraps it in the typed H3_RELAY_INTERPOLATION bundle that H3 Relay · Interpolate consumes.

    The model it loads is a standard ComfyUI frame-interpolation checkpoint - for the reference workflow that's rife_v4.26_heavy.safetensors, RIFE's heavier variant, sourced from Comfy-Org/frame_interpolation on HuggingFace. Nothing exotic: it's the same RIFE the rest of the ecosystem uses for video upscaling companions and motion smoothing.

    What you set

    Two inputs, and really only one matters:

    • model_name - dropdown populated from your models/frame_interpolation folder. Empty dropdown means you haven't dropped a checkpoint there yet.
    • manual_cache_revision (default v1) - an advanced reset key. The tooltip spells it out: change it only when you replace model contents without changing the filename. Normal model swaps are tracked automatically via fingerprint.

    How it works

    Like every H3 Relay loader, this one is a graph-builder. It expands into the core FrameInterpolationModelLoader node (which is ComfyUI's own native frame-interpolation runtime - H3 Relay deliberately reuses it rather than shipping its own) plus an internal bundle pack that stamps the result with a cache_tag fingerprint of the model name and revision.

    That fingerprint is the important bit. H3 Relay's cache keys interpolation artifacts off it, so switching interpolation models invalidates only the derived interpolated segments - your raw H3 and LTX work stays untouched. And because interpolation is chunked and encoded immediately, a model swap doesn't force you to re-do hours of generation.

    The single output, interpolation (H3_RELAY_INTERPOLATION), wires into H3 Relay · Interpolate. That's its only consumer, really.

    Install

    Shared pack install: ComfyUI Manager → search H3 Relay → install → restart, or git clone https://github.com/akatz-ai/h3-relay into custom_nodes. Then the actual model file:

    hf download Comfy-Org/frame_interpolation \
      frame_interpolation/rife_v4.26_heavy.safetensors \
      --local-dir models
    

    (from the ComfyUI root, after hf auth login). FFmpeg must be on PATH - interpolation encodes each chunk immediately, and no FFmpeg means no interpolation.

    The one gotcha beginners hit: the dropdown is empty because they cloned the pack but never downloaded the checkpoint. MODELS.md has the exact command. Interpolation model licenses are separate from H3 Relay's GPL-3.0 source code, so glance at them too.

    CategoryH3 Relay/loaders

    Inputs (2)

    NameTypeDefaultDescription
    model_nameCOMBO0 options:
    manual_cache_revisionSTRINGv1Manual cache reset only. Change this when replacing model contents without changing the filename; normal model, LoRA, dtype, strength, and attention changes are tracked automatically.

    Outputs (1)

    NameTypeDescription
    interpolationH3_RELAY_INTERPOLATION