GMFSS Fortuna VFI

If your source clip is anime - flat colors, hard line art, that stylized non-photographic look - this is the interpolation node to reach for. GMFSS Fortuna is, in its author's own words, "the all-in-one GMFSS, dedicated for anime video frame interpolation." Most VFI models were trained on real camera footage and they get twitchy on drawn content: they smear line art and hallucinate mush in the flat regions where there's no texture for optical flow to grab onto. GMFSS Fortuna was tuned for exactly that failure mode.

Frame interpolation, if you're new to it, means inventing new frames between your existing ones to make motion smoother. You feed in a choppy low-fps clip and get back a batch with 2x (or more) the frames - no re-running the diffusion model, which is the expensive part. It's the standard finishing move on AnimateDiff or Wan output, where you generate few frames to save time and then interpolate up to something watchable.

Why anime is its own problem

Real footage has grain and texture everywhere, which gives motion-estimation models something to lock onto. Anime doesn't. Big flat cel-shaded areas and crisp outlines confuse general interpolators into wobbling the lines or blending frames into a blur. AnimateDiff's whole surviving niche in 2026 is that specific hand-drawn, morphing aesthetic - the look people say "nothing else quite replicates" - and if you're chasing that, a general interpolator will fight you. GMFSS Fortuna is the one built to keep the lines clean.

How it works

GMFSS stands for "GMFlow-based Splatting Sync." It estimates motion with a flow model, then uses forward splatting (pushing pixels along the flow) plus a fusion/refinement stage to synthesize the middle frame. The "Fortuna" variant bundles the pieces together and tunes them for animation. You don't configure any of it - the model handles the interpolation internally and you just pick which weights to load.

The inputs that matter

• frames (IMAGE) - the clip you're smoothing, at least 2 frames, wired from your VAE decode or a Load Images node.
• ckpt_name - two options here: GMFSS_fortuna_union and GMFSS_fortuna. The union build is the more general one; both download automatically on first run. If you're unsure, start with union.
• multiplier (INT, default 2) - output frames per input gap. 2 doubles the frame count, 4 quadruples it. Anime holds up well at 2x; go higher only if you genuinely need buttery motion.
• clear_cache_after_n_frames (INT, default 10) - lower it if you hit out-of-memory on a long clip, raise it for a little more speed on short ones.

There's an optional optional_interpolation_states input for feeding in a Make Interpolation State List, if you want to skip interpolating across hard cuts. Usually left unwired.

Output is a single IMAGE batch - the expanded frames. Pipe it into a Video Combine node (VideoHelperSuite) to write the file.

Installing it

You install the whole ComfyUI Frame Interpolation pack:

• ComfyUI Manager - search ComfyUI Frame Interpolation, install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/Fannovel16/ComfyUI-Frame-Interpolation, then python install.py (Linux, inside your ComfyUI venv) or install.bat (Windows). Restart.

The GMFSS checkpoints pull down on first use, so budget a slow first run.

Common issues

The pack's dependency chain runs through cupy, the CUDA ops library, and that's where installs break. Windows users should run install.bat rather than install.py - the batch script handles cupy correctly. If interpolation is somehow running on CPU with the GPU idle, that's the backend having failed over, a genuinely common gripe with this pack; verify cupy installed against your actual CUDA version.

If it OOMs on a long clip, drop clear_cache_after_n_frames first. And a bigger-picture note: interpolation smooths what's already there - it can't fix bad or inconsistent source frames, so if your anime clip is already flickering, fix that upstream before you blame the interpolator. On a non-NVIDIA machine you can try the taichi backend (pip install taichi, set ops_backend: taichi in config.yaml), but expect a NotImplementedError on any node taichi doesn't cover.