AMT VFI

AMT is one of the newer, quality-leaning options in the ComfyUI Frame Interpolation pack. When your choppy clip has genuinely complex motion - multiple things moving in different directions, occlusions, stuff passing in front of other stuff - AMT tends to hold together where lighter models smear. It's the one to try when FILM or RIFE gave you ghosting and you're willing to trade some speed for a cleaner result.

Frame interpolation, quickly: it invents new frames between the ones you already have, so a low-fps clip plays back smooth. You generate video cheaply at a low frame count with something like AnimateDiff or Wan, then interpolate 2x or 4x to hit a watchable frame rate without paying to diffuse every real frame. AMT is the finishing step at the end of that chain.

How it works

AMT stands for "All-Pairs Multi-Field Transforms," from a CVPR 2023 paper. The short version: instead of estimating a single optical-flow field between two frames, it builds all-pairs correlation and derives multiple flow fields, then combines them. That redundancy is why it copes with tricky, layered motion better than single-flow methods - when one flow estimate is wrong at an occlusion boundary, the others can cover for it. You don't tune any of that; you just choose how big a model you want to run.

The inputs that matter

• frames (IMAGE) - the clip you're smoothing, at least 2 frames, straight from your VAE decode or a Load Images node.
• ckpt_name - four weights, and this is the real decision: amt-s (small, fastest), amt-l (large), amt-g (the biggest, best quality, heaviest), and gopro_amt-s, a small model fine-tuned on the GoPro dataset for real, fast camera footage. Start with amt-s; step up to amt-l or amt-g if you need more fidelity and have the VRAM. All download automatically on first run.
• multiplier (INT, default 2) - output frames per gap. 2 doubles, 4 quadruples. Start at 2.
• clear_cache_after_n_frames (INT, default 1) - note AMT ships with this defaulted to 1, not 10 like most nodes in the pack. That's a tell that AMT is memory-hungry; it clears VRAM after every frame out of the box. Raising it speeds things up but risks OOM, so nudge it carefully. Its max here is 100.

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

The output is a single IMAGE batch - the expanded frames. Feed it into a Video Combine node (VideoHelperSuite) to write your file.

Installing it

It's part of 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 AMT checkpoints download on first use.

Common issues

Out-of-memory is the thing to watch with AMT - its default clear_cache_after_n_frames of 1 is a hint that it's the memory-heavy option here. If you're pushing a long clip or the amt-g model and you OOM, keep that value low (it's already at the floor) and drop your resolution or multiplier instead. If you have headroom, raising it gets you more speed.

The pack-wide gotchas apply too: the dependency chain runs through cupy, so on Windows run install.bat (not install.py) to get it set up correctly. If interpolation is crawling with the GPU idle, the backend has likely fallen back to CPU - a recurring complaint with this pack - so verify cupy installed against your CUDA version. And remember interpolation only smooths what's there; it won't rescue frames that were already flickering upstream.