Load SPEED Model

Load SPEED Model is the loader for the pack's modern default: SPEED, a one-step pixel diffusion model for frame interpolation. It's the current hotness for good reason - diffusion-based interpolation, one generation step per midpoint, ~447MB checkpoint, and quality that competes with the non-diffusion pairwise models while being much simpler to think about. If the pack's README's TL;DR is anything to go by, try SPEED first and only reach for the others when you hit its limits.

Two quirks make this loader worth reading about before you run it:

• It downloads two things. The speed.pt checkpoint (~447MB) comes from the official SPEED HuggingFace repo (zhZ524/SPEED) into ComfyUI/models/speed-vfi/. But SPEED's inference code isn't something Tween bundles - so on first use the loader also fetches a checksum-pinned snapshot of the official runtime (the source archive at a specific commit, SHA-256 verified) and unpacks it next to the checkpoint. That's why first load is slower than the file sizes suggest, and why you might see an unexpected zip download happen. It's also why the license is worth a glance: the upstream SPEED repo had no license file when this was pinned, and Tween deliberately doesn't redistribute that source - it pulls it from upstream directly.
• precision is a real choice. Options are auto (default), bf16, fp16, fp32. Auto picks BF16 on supported CUDA GPUs (Ampere+), otherwise FP16 - the sensible default. FP32 is there for comparison or debugging: slower, more VRAM, and almost never better to the eye.

The model_path dropdown just lists checkpoints in models/speed-vfi/, defaulting to speed.pt. Output is a model wire into the SPEED Interpolate / Segment Interpolate nodes. The loader also handles the model's 64-pixel input padding requirement - inputs get padded to the divisor and cropped back automatically, so you don't have to think about resolution math.

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. Unlike the Google Drive–backed loaders, SPEED pulls from HuggingFace, which is generally faster and more resumable. First run does the checkpoint and the runtime snapshot.

Gotchas. Don't mistake the first-run runtime download for a hang. If you're on an older GPU without BF16 support, auto still picks the right thing, so leave it alone. And SPEED is stochastic - the model's output depends on a seed, which is set on the Interpolate node, not here - so don't be puzzled when two runs don't match.