Load GIMM-VFI Model

Load GIMM-VFI Model is the entry point for the pack's "arbitrary timesteps, efficient 4x/8x" model. GIMM-VFI (NeurIPS 2024, S-Lab NTU) works differently from the pairwise crowd: instead of only knowing how to make a midpoint, it learns an implicit motion model that can generate frames at any timestep between two input frames in a single pass. That's what lets the Interpolate node do 4x or 8x without three recursive passes.

Two things make this loader interesting:

• It pulls two models, not one. model_path picks the GIMM checkpoint - gimmvfi_r_arb_lpips_fp32.safetensors (RAFT flow variant, ~80M params) is the default; gimmvfi_f_arb_lpips_fp32.safetensors (FlowFormer variant, ~123M) is the higher-accuracy option. But GIMM-VFI needs an optical flow estimator to drive its motion model, so the loader also auto-downloads the matching one - raft-things_fp32.safetensors for the RAFT variant, flowformer_sintel_fp32.safetensors for FlowFormer - into ComfyUI/models/gimm-vfi/ alongside it. First run is a two-file download from HuggingFace, so it's slower to start than it looks; subsequent runs are instant.
• ds_factor is the VRAM dial. It's a downscale factor for internal processing: 1.0 = full resolution, 0.5 = half. Lower it and everything - flow estimation, motion modeling, synthesis - runs at reduced resolution internally, cutting VRAM and speeding things up at the cost of fine detail. The tooltip's advice is the one to follow: try 0.5 for 4K inputs. GIMM is ~2.5GB per pair at full res, and 4K can push that past comfort on mid-range cards.

It's a port of Kijai's ComfyUI-GIMM-VFI adaptation (the pack credits him and uses his safetensor checkpoints), so if you've seen GIMM-VFI in that ecosystem, this is the same model with Tween's unified controls on top. Output is a model wire into the GIMM-VFI Interpolate / Segment Interpolate nodes.

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, then let the loader handle the downloads on first execution. Like BIM and SGM, GIMM gets a speed bump from cupy (its warp falls back to PyTorch if cupy isn't installed), so the cupy-cudaXXx matching your torch build is worth it.

Gotchas. First load takes a while because it's fetching the GIMM checkpoint and a flow estimator - don't kill it. If you switch variants, expect the second flow model to download too. And remember ds_factor is a quality/VRAM trade: crank it down for 4K, leave it at 1.0 for anything you care about.