Flow Visualizer (Warper)

Every video workflow eventually asks "is this motion actually smooth, or is it two frames of chaos?" Flow Visualizer (Warper) answers with pictures: it runs an optical-flow model over your frame sequence and paints each frame's motion as a color-coded image - hue encodes direction, brightness encodes speed. Wire it into a preview and you can literally see whether the camera pans, the subject drifts, or the motion breaks, before you waste a sampling run. It's the diagnostic tool the video side of ComfyUI quietly lacks.

It's the model-dependent node in ComfyUI Warper Nodes, the niche pack by workflow author AIWarper. Unlike the rest of the pack, this one has a real model download and real dependencies, and it's the only node in the pack that will fail outright if you skip the setup. Worth it if you do any serious video or frame-interpolation work.

How it works

The pack bundles SEA-RAFT (a modern RAFT-family optical-flow model, shipped in the repo's sea_raft folder). The node takes your images batch, computes flow between each frame and the next (frame i against i+1, wrapping), pads dimensions to a multiple of 8 for the model, and runs it on CUDA. Output depends on direction: "forward", "backward", or "both" (default). The raw flow vectors get converted to the standard color-wheel visualization via flow_to_color, and each direction's frames come out as an IMAGE you can preview or save.

The one input that trips everyone up is checkpoint - it's not a path you type, it's a dropdown populated from whatever .pth files sit in ComfyUI/models/raft/. Empty dropdown means no model placed. direction controls which flows you get; "both" produces both outputs, "forward" gives you the forward output and a black frame in the backward slot.

Installing it (this is the one with a model)

Everything else in this pack is clone-and-go; this node is not. First install the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/AIWarper/ComfyUI-WarperNodes

Then get the RAFT checkpoint. The README points at a Google Drive folder; grab Tartan-C-T-TSKH-spring540x960-M.pth and put it at:

# create the folder if it doesn't exist
mkdir -p ComfyUI/models/raft
# place Tartan-C-T-TSKH-spring540x960-M.pth inside ComfyUI/models/raft/

Finally, the pack's only Python requirements live in sea_raft/requirements.txt (torch, torchvision, numpy, opencv-python, scipy, einops, tqdm and friends). Most will already be present, but run it if you hit import errors:

cd ComfyUI/custom_nodes/ComfyUI-WarperNodes
pip install -r sea_raft/requirements.txt

Then restart ComfyUI. If the checkpoint dropdown is empty after all that, the model isn't where the node is looking.

Common issues

This node requires a GPU - the flow model runs on .cuda() unconditionally, so CPU-only setups get a hard error, not a graceful fallback. That's the biggest practical gotcha. Second: no model file, empty dropdown, and if you wire a name in manually it'll throw a "model not found" - double-check the models/raft/ path matches your ComfyUI install. Expect it to be slower than a preview node has any right to be: RAFT is a real model running per-frame-pair in both directions on "both". And if the flow viz looks like static, your frames may genuinely be low-motion (a static shot produces near-black flow) - that's correct output, not a bug. It's a diagnostic, so read it like one.