GIMM-VFI Interpolate

Every other pairwise interpolator in this pack does 4x or 8x by running the same 2x trick two or three times, recursively. GIMM-VFI Interpolate is the one that says "why?" - it generates all the intermediate frames for a pair in a single forward pass, because the model (NeurIPS 2024) learns an implicit motion model that can answer for any timestep, not just the midpoint. That's the single_pass toggle, and it's on by default for a reason: 4x and 8x get meaningfully faster, and the frames are generated together rather than compounded through error-prone recursive passes.

Inputs. The familiar shared set - images, model (from Load GIMM-VFI Model), multiplier, batch_size, chunk_size, keep_device, all_on_gpu, clear_cache_after_n_frames, source_fps/target_fps - plus the node's signature input:

• single_pass (on by default) - the arbitrary-timestep mode that makes all intermediates in one pass. Turn it off and GIMM falls back to the standard recursive 2x approach, same as BIM/EMA/SGM. There are occasional edge cases where recursive is preferable, but you'll know them when you hit them; default is right.
• batch_size - a subtle one: it's ignored in single-pass mode, because each pair produces multiple frames and pairs are processed one at a time. In recursive mode it works like the other nodes. Don't be confused when raising it does nothing on default settings.
• multiplier - 2x/4x/8x; in single-pass mode these are one pass per pair rather than recursive.
• source_fps / target_fps - exact-FPS mode, oversample-then-select, same as the rest of the pack.

Quality-wise, GIMM is the pack's efficient arbitrary-timestep option (~2.5GB per pair), with one honest caveat: it's still frame-pair-centric, so it doesn't give you the long-range temporal coherence that the sequence-native LDF-VFI does. For the common "24 to 48/96" jobs it's a strong default and a genuinely nice speed story at 8x.

Outputs: images, oversampled (same as images in multiplier mode), elapsed_seconds.

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, let the loader fetch the GIMM checkpoint plus its flow estimator (first load is a double download), and wire the output into VHS Video Combine. cupy speeds GIMM up if you install the matching wheel.

Where people get burned: the first-run double download (it's not frozen, it's fetching a flow model), and assuming batch_size is broken because it doesn't respond in single-pass mode. For 4K inputs, drop ds_factor to 0.5 at the loader before reaching for bigger VRAM. And if frames look less temporally stable than you'd like on long sequences, that's the pair-centric design - not a bug.