Motion Ref Settings (Custom)

Motion Ref Settings (Custom) is the same idea as the pack's default settings node, but with the training wheels off. Where the default gives you three knobs, this one hands you 30-odd switches and says "good luck." If you're reaching for it, you already know the default preset isn't transferring the motion you want - too weak, too strong, or dragging the reference's subject into your frames. This is where you actually experiment.

It's part of the same experimental MotionThief pack (SD 1.5/AnimateDiff only, author's own disclaimer: hacky, for experiments, not compatible with a lot of things). Like its sibling, it doesn't do anything by itself - it builds a MOTION_REF_SETTINGS object for Apply Ref Motion.

What all the switches do

Every temporal attention layer in the AnimateDiff UNet lives in one of 8 input blocks or 12 output blocks. That's why the inputs are named the way they are:

• input_1 … input_8 - toggle each of the 8 input temporal blocks on or off.
• output_1 … output_12 - same for the 12 output blocks.

A layer that's on participates in the reference transfer during Apply Ref Motion's WRITE and READ passes. A layer that's off just runs normally. The default settings node is exactly one configuration of this one: norm_bank on, the first strength output toggles on, everything else off. So anything you can do with the default node, you can do here with finer control - plus block-level precision.

Then there are four bank toggles, which choose what gets recorded from the reference and replayed into your generation:

• k_bank / v_bank - store the reference's attention keys/values and concatenate them into your attention context. This is the "attend to the reference's motion" mode, the soft, blendable option.
• q_bank - replace your attention query with the reference's. Harder replacement, more aggressive copying.
• norm_bank - store the reference's hidden states and let your attention read from them. This is the mode the default node uses.
• norm_fidelity (float, 0–1, default 0) - blends between the reference-steered result and a re-run of your own generation's attention. Only kicks in when norm_bank is on; at the default 0 the bank simply steers via concatenation, and as you raise it, more of your own motion gets mixed back in.

Reading the wrapper code, the k/v banks concatenate into your context (steer), while q/k/v replacement swaps your tensors wholesale (much stronger, and a good way to overfit the reference clip). Start with norm_bank plus a modest norm_fidelity, and only reach for q/k/v replacement if you're deliberately chasing the reference's exact motion.

enabled (default true) and prev_settings (optional) work exactly like the default node: disable to pass the chain through untouched, chain multiple settings nodes to stack configurations.

The output

MOTION_REF_SETTINGS → into Apply Ref Motion's ref_settings input. Same as always; that's the only wire.

Installing it

ComfyUI Manager: search "ComfyUI-MotionThiefExperiment" and install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/logtd/ComfyUI-MotionThiefExperiment

No pip deps, no model downloads from the pack. The surrounding workflow needs ComfyUI-AnimateDiff-Evolved, VideoHelperSuite, an AnimateDiff motion module, and an SD 1.5 checkpoint - without a motion module applied first, there are no temporal blocks and none of these toggles mean anything.

Gotchas

• More layers ≠ better. Each layer you enable is one more place the reference's features can drag your output toward its content. If the subject's identity is leaking into your frames, you're over-selecting.
• The q/k/v toggles are the sharp edge. Full tensor replacement is powerful and unstable; if things start warping, disable them before you blame the sampler.
• There's no saved-preset mechanism here. Keep notes on which combo worked - a stack of toggles is easy to lose track of between test runs.

It's a fiddly node for a fiddly pack, and that's the point. When the default's three knobs can't get the motion right, this is where you'll find out whether MotionThief can do what you want at all.