Emiewn GIMM-VFI Interpolate

This is the node that does the actual work in the Emiewn GIMM-VFI pair. The Load node gets you a model; this one takes a batch of frames and produces the in-between ones - so a 16fps video becomes 32fps, or a clip gets the extra frames to slow down smoothly without turning into a stuttery slideshow. If you've ever generated a video in ComfyUI (Wan 2.2 or otherwise), seen the motion judder, and wondered how people get that buttery 60fps look, frame interpolation is the answer, and GIMM-VFI is one of the best interpolators available locally.

How it works

For each adjacent pair of frames it estimates optical flow (using the flow model your Load node picked - RAFT for the R variant, FlowFormer for the F), then uses that motion to predict what the scene looked like at intermediate moments. interpolation_factor controls how many in-betweens: at 2, each pair becomes three frames (I0, mid, I1); at 4, it's five. It stitches the whole sequence together in order, so feeding it a batch of N frames returns N + (N-1)×(factor-1) frames. Internally it pads frames to a multiple of 32 before the model runs and un-pads the output.

Inputs that matter

• gimmvfi_model - the GIMMVIF_MODEL from Emiewn Load GIMM-VFI. Nothing to set; just wire it.
• images (IMAGE) - the frame batch to interpolate between. Extract frames from a video or a generated clip first, in order.
• interpolation_factor (integer, 1–100, default 2) - the multiplier. 2 is the sane default; 4+ gets expensive fast and can smear on complex motion.
• ds_factor (float, 0.01–1, default 1) - the model's internal downsampling ratio for the interpolation. Leave at 1 unless you're chasing speed on a big batch.
• seed (integer) - seeds the motion sampling so runs are reproducible.
• output_flows (boolean, off) - set this and the node also emits color-coded optical-flow visualizations, which is genuinely useful for debugging "why is the interpolation warping here."

Outputs: images (IMAGE, the full interpolated sequence) and flow_tensors (IMAGE - the flow visualizations when enabled, or a dummy 1×64×64 tensor when not).

Installing it

The two GIMM-VFI nodes share one install, and the hard part is the sibling dependency. You need Kijai's ComfyUI-GIMM-VFI in custom_nodes because this node reuses its model configs and utilities:

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 (or install both through ComfyUI Manager). Models auto-download to ComfyUI/models/interpolation/gimm-vfi/ on first use.

Performance and expectations - be honest with yourself

GIMM-VFI is a quality-first tool, not a speed tool. The community numbers are blunt: one user compared 2× interpolation on 81 frames and RIFE took ~50 seconds while GIMM took ~4 minutes - roughly four times slower - and the trade is better handling of fast motion and complex scenes. The F model is slower than the R model but is the one people call "the highest quality frame interpolation model I've been able to find."

Where people get burned:

• Interpolation factor 8 or 16 on a long clip will test your patience and your VRAM. Start at 2.
• Heavy motion with occlusions (things moving behind other things) is where any interpolator - GIMM included - produces warping. That's what output_flows is for: eyeball the flow and see if the motion estimate went wrong.
• This pack exists because the author's official GIMM-VFI install broke. If yours works, the Kijai original is the more battle-tested path; this is the fallback that skips the cupy build.