DWposeDeluxe Reverse Keypoints

This one's almost too simple to need explaining: feed it a POSE_KEYPOINT batch and it reverses the temporal order of the frames, then hands the flipped sequence back. The end becomes the beginning; the beginning becomes the end. That's the entire job, and it's genuinely handy more often than you'd think.

Why you'd want a reversed pose

The most common use is loop construction. Video models love a seamless cycle, and a pose loop is the cheapest way to get one: take any motion clip, reverse it, and append the reversed copy to the original with BatchKeypoints. Forward motion, then backward motion, forever. You can do exactly the same thing to close a PoseInterpolation loop - interpolate A→B, reverse it, splice, and the cycle is seamless.

Reversing poses is also a neat debugging trick. If you suspect temporal corruption in a dataset - a frame that got dropped or reordered somewhere in your pipeline - reverse the batch, run it through KeypointDiff against the correctly-ordered original (reversed), and the diff will pinpoint where the sequence stops matching.

The one thing to keep straight

This reverses time, not space. It does not mirror the pose left-to-right, and it does not flip the coordinates. If what you actually want is a mirrored dancer - a left-handed move turned right-handed - this is the wrong node; you'd need a horizontal flip on the x-coordinates instead (which PoseResize's offset tools don't do either). Don't confuse the two, or your "reversed" dance will look like a person walking backwards rather than a mirrored one.

Inputs and output

Single required input pose_keypoints; single output pose_keypoints with the frames in reverse order. Nothing else to configure, which is refreshing after the pack's more parameter-heavy nodes.

Installing it

Part of the DWposeDeluxe pack:

cd ComfyUI/custom_nodes/
git clone https://github.com/hobinrude/ComfyUI_DWposeDeluxe
cd ComfyUI_DWposeDeluxe
pip install -r requirements.txt

or search DWposeDeluxe in ComfyUI Manager. Pure list reversal, no models, instant on any batch size. Since it's such a simple op, the main failure mode is feeding it a single dict or JSON string instead of a list - the pack's nodes normalize that for you, but a non-empty-but-wrong structure will just produce a warning and empty output. Check the console if you get nothing back.