(Down)Load GIMMVFI Model

If you've already fought with RIFE VFI and it still smears your fast-motion clips into soup, this is the node that fixes it. (Down)Load GIMMVFI Model is the loader half of kijai's GIMM-VFI pack - it fetches the model weights and hands them off to GIMM-VFI Interpolate, the node that actually does the frame-making. You don't do anything creative here; you just pick which model and how much precision you're willing to trade away, and this node gets it onto your GPU.

Why GIMM-VFI over the usual RIFE/FILM combo

GIMM stands for Generalizable Implicit Motion Modeling, and it's a different research lineage than the RIFE/FILM/GMFSS pile that ships in ComfyUI's built-in Frame Interpolation pack. The pitch, and what people actually report using it: it holds together on motion that makes RIFE fall apart - fast pans, chaotic action, things crossing behind other things. The trade-off is speed. Community timing on the same 480p clip put RIFE at 9 seconds, FILM at 12, and GIMM's two variants at 25 and 53 seconds. That's not a rounding error, that's five-to-six times slower. So this isn't your default interpolator - it's the one you reach for on the specific shot that RIFE botched.

The one input that actually matters

model is a dropdown with exactly two choices, and the choice matters more than anything else on this node:

• gimmvfi_r_arb_lpips_fp32.safetensors - the "R" variant. Faster.
• gimmvfi_f_arb_lpips_fp32.safetensors - the "F" variant. Slower, and the one people report as noticeably better once there's real motion in frame.

If you're not sure which to grab, start with R - it's roughly half the runtime of F in reported timings, and for a lot of clips the quality gap won't matter. Reach for F when the R output looks warped or ghosted on a specific hard shot.

The two optional settings are lower-stakes. precision defaults to fp32; bf16 and fp16 are also on offer if you're tight on VRAM, but this is research code, not a production-hardened quantized checkpoint, so don't be shocked if lower precision introduces artifacts the fp32 default doesn't have. torch_compile is off by default - flipping it on can speed up repeated runs but adds real compile time to your first execution, so it's only worth it if you're going to run the same model shape a lot in one session.

The single output, gimmvfi_model, is a GIMMVFI_MODEL object. It goes straight into the gimmvfi_model input on GIMM-VFI Interpolate - nothing else in ComfyUI consumes this type, so that's its only destination.

One small landmine when you're hunting for this node in the menu: its category still reads PyramidFlowWrapper, a copy-paste leftover from another kijai pack. It's not miscategorized garbage, it's just a label nobody fixed - search by node name, not by browsing that category.

Installing it

Two paths, same as most kijai packs: ComfyUI Manager, search "ComfyUI-GIMM-VFI," install, restart. Or manually: cd ComfyUI/custom_nodes && git clone https://github.com/kijai/ComfyUI-GIMM-VFI, then restart Comfy.

The README is short but the one line that matters is load-bearing: it requires cupy, tested specifically against cupy-cuda12==13.3.0. This isn't optional glue - GIMM-VFI's CUDA ops depend on it, and unlike the RIFE/FILM pack, there's no documented CPU or non-CUDA fallback path here. If you're on a card without a matching CUDA toolkit, or on Apple silicon, expect this pack to just not work, full stop.

The model file itself isn't bundled - pick your model value in the node, run it once, and the safetensors file downloads automatically into ComfyUI/models/interpolation/gimm-vfi/. That first run will be slower while it fetches; after that it loads from disk.

Where people get stuck

cupy version mismatches. Since the README pins a specific cupy-cuda12 build, a cupy install that doesn't match your actual CUDA version is the most common way this pack fails to import at all. If ComfyUI throws an import error on startup pointing at this node pack, check your cupy version against your CUDA toolkit before assuming the node itself is broken.

Picking F "because it's better" and then waiting forever. It's tempting to default to the higher-quality variant, but on long clips that 2x+ time penalty adds up fast. Use R as your default and only switch to F on the shot that actually needs it.

Confusing this for the RIFE-style pack. If you already have ComfyUI-Frame-Interpolation installed, you'll have RIFE, FILM, and friends in your node list too. (Down)Load GIMMVFI Model and GIMM-VFI Interpolate are a separate pack entirely - installing one doesn't touch the other, and you need both nodes from this pack together for GIMM-VFI to do anything.