ATM VFI (only supports 2x multiplier)

ATM is the newest model in the ComfyUI Frame Interpolation pack - a 2025 transformer-based interpolator, so if you want to try the current state of the art rather than a battle-tested 2022 model, this is the one to poke at. The tradeoff is the usual for the freshest option: it only does 2x interpolation, and it's not the fastest thing in the pack.

Frame interpolation, in a sentence: it invents new frames between your existing ones so a choppy low-fps clip plays smooth. You generate video cheaply at a low frame count with AnimateDiff or Wan, then interpolate up to a watchable frame rate instead of paying to diffuse every real frame. ATM is one option for that step, and a modern one.

How it works

ATM stands for "Attention-to-Motion," from a 2025 ICASSP paper. Instead of leaning purely on convolutional flow estimation, it uses a transformer to relate motion across the frame via attention - the idea being that attention can pick up long-range and global motion (a whole camera pan, say) that a local convolutional model handles less gracefully. The global_motion control below is the user-facing handle on exactly that capability. As with the rest of the pack, you don't touch the internals; you choose a checkpoint and a motion mode.

The inputs that matter

• frames (IMAGE) - your clip, at least 2 frames, from a VAE decode or a Load Images node.
• ckpt_name - three weights: atm-vfi-base.pt (the standard), atm-vfi-lite.pt (lighter and faster), and atm-vfi-base-pct.pt (a base variant with different tuning). Start with base, drop to lite if it's too slow or heavy. All download on first run.
• multiplier (INT, default 2) - locked at 2. ATM only does 2x, as the node's own display name spells out. For a bigger jump, run a second interpolation pass or use a higher-multiplier model.
• global_motion - three modes: On, On with Ensemble (slowest), and Off (fastest). This toggles ATM's global-motion handling. On is the sensible default for footage with camera movement; On with Ensemble runs extra passes for a quality bump at a real speed cost; Off is fastest and fine when the camera is mostly static. If a pan looks smeary, this is the first knob to turn.
• clear_cache_after_n_frames (INT, default 10) - the OOM valve; lower it on long clips if memory gets tight.

An optional optional_interpolation_states input takes a Make Interpolation State List for skipping specific frame pairs. Usually left unwired.

Output is a single IMAGE batch - send it into a Video Combine node (VideoHelperSuite) to encode.

Installing it

It ships in 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, in your ComfyUI venv) or install.bat (Windows). Restart.

The ATM checkpoints download on first use. As a newer addition, make sure your pack install is up to date (git pull in the custom node folder, or update via ComfyUI Manager) - an older copy may not have this node at all.

Common issues

If ATM feels slow, that's partly the transformer design and partly your settings: switch global_motion from On with Ensemble to plain On or Off, and try the lite checkpoint. Those two changes buy back most of the speed.

The pack-wide install issues still apply: cupy is the dependency that breaks installs, so run install.bat on Windows rather than install.py. A GPU sitting idle during interpolation means the backend fell back to CPU - check cupy against your CUDA version. And since ATM is 2x-only, plan a second pass if you need higher frame-rate multiplication.