Emiewn Load GIMM-VFI

Frame interpolation in ComfyUI has a long history of "just works" nodes that are mediocre, and great nodes that are a pain to install. GIMM-VFI is one of the great ones - an optical-flow-based interpolator that handles fast, complex motion better than the usual RIFE default - and the pain in question is usually cupy, the compiled CUDA dependency its softsplat implementation needs. This node is one author's "fine, I'll do it myself" answer: load a GIMM-VFI model with a built-in pure PyTorch softsplat, no cupy required.

The README is refreshingly honest about the motivation: "Only reason for GIMM-VFI nodes is that I previously had some issues with the official nodes not working so I needed an alternative." That's the whole story in one sentence. If Kijai's official GIMM-VFI nodes already work for you, you don't need this. If they don't - and on Windows, cupy builds are the usual culprit - this is the fallback.

How it works

Under the hood it imports the actual GIMM-VFI code (GIMMVFI_R/GIMMVFI_F model classes, RAFT and FlowFormer optical-flow estimators) and monkey-patches the cupy-dependent softsplat module with a pure PyTorch reimplementation before anything imports it. Same math, no compiled extension. On first load it downloads the weights from Kijai's HuggingFace repo to ComfyUI/models/interpolation/gimm-vfi/ - both the interpolation model and the optical-flow model it pairs with.

Inputs

• model - pick one of two:
  • gimmvfi_r_arb_lpips_fp32.safetensors - the R model, RAFT-based flow. Faster, the everyday choice.
  • gimmvfi_f_arb_lpips_fp32.safetensors - the F model, FlowFormer-based flow. Noticeably slower, higher quality. The community consensus is that GIMM-VFI-F is about the best frame interpolation available locally.
• precision (fp32 default, bf16, fp16) - fp32 is safest. Drop to bf16 or fp16 if VRAM is tight; quality loss is usually minor.
• torch_compile (boolean, off by default) - compiles part of the model for speed. Only works where Triton is available, which mostly means Linux; on Windows this is where things get flaky. Leave it off until you've confirmed it runs.

Output is a single gimmvfi_model (GIMMVIF_MODEL), which feeds into Emiewn GIMM-VFI Interpolate in the same pack.

Installing it - read this part

This is one of two nodes in the pack that has a hard dependency beyond the pack itself. You need Kijai's ComfyUI-GIMM-VFI installed as a sibling folder in custom_nodes - the README says so explicitly, because this node borrows its model configs and utility imports from there:

cd ComfyUI/custom_nodes
git clone https://github.com/emiewnn/ComfyUI-Emiewn-Nodes.git
git clone https://github.com/kijai/ComfyUI-GIMM-VFI.git

Restart ComfyUI after both. (Or install both via ComfyUI Manager - search "ComfyUI-Emiewn-Nodes" and "ComfyUI-GIMM-VFI".) The pack's own requirements - huggingface-hub, opencv-python, pyyaml, omegaconf - install automatically.

Gotchas

• First run downloads several GB (fp32 weights plus the flow model). Let it finish before judging anything.
• No requirements.txt fight with cupy - that's the whole point. If your GIMM-VFI adventures died at "no module named cupy" before, this sidesteps it.
• You still need the interpolation node. Loading a model does nothing by itself; wire it into Emiewn GIMM-VFI Interpolate and feed it frames.