Load LDF-VFI Model

Every other model in the Tween pack is a frame-pair interpolator. Load LDF-VFI Model loads the one that isn't: LDF-VFI, a holistic sequence-diffusion transformer that models your whole clip's temporal structure instead of one pair at a time. That's what makes its output temporally coherent in a way pairwise nodes can't touch - no flicker, no drift, no seams - and it's also why it's the pack's heavyweight. Read the numbers before you commit: ~6.4GB of weights, roughly 20GB VRAM to run, and it's slow. This is a "sequence consistency matters more than speed or memory" tool, per the pack's own framing.

What the loader sets up. The model dropdown defaults to onecat-ai/LDF-VFI (the official HuggingFace repo) - on first use it downloads the transformer and a conditional VAE (~6.4GB total) into ComfyUI/models/ldf-vfi/, plus a checksum-pinned snapshot of the official Apache-2.0 runtime, the same pattern as SPEED. Loading stays on CPU until the interpolation node actually runs, so the loader itself is cheap; the VRAM cost lands when you execute.

The other inputs are the VAE tiling knobs, and they're the difference between "fits" and OOM:

• tile_size (256) / tile_overlap (64) - spatial tiling for the VAE, with seam blending across the overlap. Larger tiles can be more efficient; the defaults are the sane start.
• vae_batch_size (8) - temporal-tile batch size. Lower this first whenever VAE encode/decode OOMs.
• attention_type - defaults to the official slide_chunk_all_block_2x1x1 sparse attention, which the README recommends. Full attention is theoretically better and practically a memory bomb - leave it alone.

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

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

LDF-VFI has extra requirements the other nodes don't: PyTorch 2.5+ and a current diffusers/accelerate stack. The pack's requirements.txt pins diffusers>=0.33.1,<0.40 and accelerate>=1.5,<2, which Manager installs for you - but if your ComfyUI venv runs an older torch, upgrade it before blaming the node. First loader execution is a multi-gigabyte download; go make coffee.

Gotchas. The 20GB VRAM figure is the official 8x quick-start at benchmark resolutions - higher resolutions may need smaller VAE tiles or more VRAM. If you're on a 12GB card, this node isn't for you, full stop. And because loading stays on CPU until execution, a node that looks "stuck" at load time may just be waiting for the interpolate step. There's no Segment variant of LDF for a reason - splitting the sequence externally would throw away the long-range context this model exists to preserve.