Wan Video Frames Padding

Wan is a diffusion transformer with a very specific appetite: it wants its frame counts in 4*n + 1 shape. 81, 121, 161 - fine. 83 or 122 - it doesn't care how close you think you are, that's a broken run or a mysterious error. This node is the fixer: give it a batch of frames, and it pads to the next valid count by repeating the last frame.

That's the entire job, and it's a real one. Any Wan workflow that sources frames from outside - an input video with an odd length, a frame sequence you built yourself, a face-swap or interpolation pipeline that spits out arbitrary counts - will eventually hit this wall. The standard solution in the wild is a hand-rolled pad node or a "round to multiple" hack; this one is purpose-built and comes with the rest of the Turing Utils pack.

How it works

The mechanism is the least glamorous thing in this pack: copy the final frame N times and append. The node's own _ceil_wan_frame_count rounds frame_count up to the next 4*n+1. If you give it an explicit target_frame_count instead, it pads exactly to that - but only if that count is itself 4*n+1, and only if it isn't shorter than the input. This node pads; it never trims. Trim upstream if your video is too long.

The inputs that matter

• image - the IMAGE batch [frames, height, width, channels].
• target_frame_count - 0 (default) means "round me up to the next valid count." Set a specific number if you need an exact length.
• mask (optional) - padded with the same repeated last frame, so a mask tracks the video frame-for-frame.

Outputs: image, mask, plus width, height, length (the padded output length), and input_length (what went in) - the last two are handy for wiring into a text or logging node to see how much padding actually happened.

How to install

It ships with ComfyUI Turing Utils ("comfyui-svdint4" in Manager):

cd ComfyUI/custom_nodes
git clone https://github.com/wjie98/comfyui-svdint4
cd comfyui-svdint4
python -m pip install -v --no-build-isolation -e ./kernel

The kernel build step is manual for the pack as a whole, though this particular node is pure tensor logic - it works even if you've only installed the Python side. Restart ComfyUI either way.

The honest take

The repeat-last-frame trick is the community-standard way to pad video for Wan, and it's the right call for short gaps: one or two repeated frames are invisible to the model. It stops being invisible if you're padding dozens of frames, so don't use this node to stretch a 30-frame clip to 81 and expect magic - at that point the repeated frames become a frozen tail the model has to reconcile. For the "my input is 79 frames, Wan wants 81" problem, it's the one you reach for, and it's over in seconds.