ComfyUI Node

Load EMA-VFI Model

The EMA-VFI loader where the checkpoint filename secretly decides your model's behavior

By ethanfel·Created 7 months ago·Updated about a month ago· 17
Load EMA-VFI Model
    • model
    model_pathours_t.pkl
    ttafalse

    Load EMA-VFI Model is the entry point for the pack's speed-focused frame interpolator. It's a simple loader - download the checkpoint, hand you a model object for the EMA-VFI Interpolate nodes - but there's one genuinely confusing bit hiding in the model_path dropdown: the filename encodes what the model can do. EMA-VFI ships four official checkpoints, and Tween auto-detects both the variant and its timestep support from the filename:

    | Checkpoint | Variant | Params | Arbitrary timestep | |---|---|---|---| | ours_t.pkl | Large | ~65M | Yes | | ours.pkl | Large | ~65M | No (fixed 0.5 midpoint) | | ours_small_t.pkl | Small | ~14M | Yes | | ours_small.pkl | Small | ~14M | No (fixed 0.5) |

    Default is ours_t.pkl - the large, flexible one - which is the right call for almost everyone. The _t suffix means the model can interpolate at arbitrary timesteps (which is what enables the node's target_fps exact-FPS selection and GIMM-style multi-frame thinking); the plain versions always assume the midpoint. The small variants are there if you're VRAM-starved, and you can feel the difference in quality.

    The inputs that matter:

    • model_path - the dropdown above. The file auto-downloads from Google Drive to ComfyUI/models/ema-vfi/ on first use. If you add other checkpoints to that folder, they appear here.
    • tta (test-time augmentation, off by default) - flips the input horizontally, runs the model on both, and averages the results. About 2x slower for a small quality bump; the tooltip's advice is sound - only bother with it on the large model.

    Output is a single model wire into EMA-VFI Interpolate or Segment Interpolate.

    Install. Search "Tween" in ComfyUI Manager, or:

    cd ComfyUI/custom_nodes
    git clone https://github.com/Ethanfel/ComfyUI-Tween
    pip install -r requirements.txt
    

    Restart, add the node, and the download kicks off on first execution. No manual model wrangling needed.

    Gotchas. First-run stalls are the Google Drive fetch, not a hung node - give it a minute. If you switch checkpoints and the behavior seems to change beyond just quality, that's the timestep support flipping between _t and non-_t; when a fixed-0.5 model meets a target_fps workflow the interpolation gets dumber, not broken, which is easy to misread as a bug. And unlike SGM or GIMM, EMA-VFI doesn't use cupy at all, so if you installed it for this pack, it's only helping the other models.

    Categoryvideo/EMA-VFI

    Inputs (2)

    NameTypeDefaultDescription
    model_pathCOMBOours_t.pklCheckpoint file from models/ema-vfi/. Auto-downloads on first use if missing. Variant (large/small) and timestep support are auto-detected from filename.
    ttaBOOLEANfalseTest-time augmentation: flip input and average with unflipped result. ~2x slower but slightly better quality. Recommended for large model only.

    Outputs (1)

    NameTypeDescription
    modelEMA_VFI_MODEL