TTM Latent Add

TTM Latent Add doesn't generate anything, and that's the point. In a TimeToMove workflow, the sampler never talks to the reference video directly - it only sees a latent with a bunch of extra data stuffed inside. This node is what stuffs it in.

TimeToMove (TTM), for context: you give Wan 2.2 a crude driving video of the motion you want (a cutout of your subject puppeteered through a scene in After Effects or whatever), plus a mask of the subject, and Wan re-renders it as a smooth photoreal clip that tracks that motion. This is the middle node of GiusTex's native-ComfyUI port of Kijai's WanVideoWrapper TTM nodes. It takes the reference latents from Encode WanVideo, a mask, and a step window, and bundles all of it into the latent dict that TimeToMove Guider will later read.

How it works

Mechanically it's a packer. It copies the reference latents into the latent dict under ttm_reference_latents, downscales your mask to match the latent resolution (by a factor of 8, or 16 if the reference latent has 48 channels - the 16x VAE variant), and stores the ttm_start_step/ttm_end_step window. When the guider runs, it reads those keys and injects the reference during denoising. That's the entire data path: encode → pack → guide.

Inputs and outputs

All four are required, plus a fifth:

• latent - the base generation latent (in the example workflow it comes from a first-frame/last-frame node). This is the latent that actually gets sampled.
• reference_latents - the output of Encode WanVideo. "Reference image to encode," per the tooltip.
• ttm_start_step (default 0) - first step where the reference gets injected.
• ttm_end_step (default 3) - the step where injection stops. This is the knob that decides how strongly the reference's motion dictates the output: a wider window means the whole sequence hews to the driving video; a narrow one is a light touch.
• ref_masks - the motion mask, and notice it's required, not optional. You need an actual mask covering the subject (the example workflow loads a mask video and runs it through an ImageToMask node). No mask, no TTM - don't leave it dangling.

Output: one latent, which splits to both TimeToMove Guider's latent input and the sampler's latent input.

Installing

Same as its siblings - zero third-party deps in this pack:

cd ComfyUI/custom_nodes
git clone https://github.com/GiusTex/ComfyUI-Wan-TimeToMove.git

Restart ComfyUI and the nodes appear under "Wan22 TimeToMove" (ComfyUI Manager works too, by pack title). You supply the Wan 2.2 model, VAE and text encoder; the example workflow additionally leans on ComfyUI-MoreEfficientSamplers for the CFG scheduler and Kijai's WanVideoWrapper for the first/last-frame node, but none of that is needed just to run this node.

Where people get burned

• Forgetting ref_masks. It's in the required list, so the node won't even validate without it, but it's the easiest thing to overlook when you copy a workflow. The mask defines what region gets animated; get it wrong and the reference bleeds across the whole frame.
• ttm_end_step below ttm_start_step. The node raises a hard ValueError telling you exactly that. It's a sanity check, not a mystery crash.
• Treating this as the sampler input. The latent this outputs is also what you feed the sampler - the reference data rides along invisibly. If you wire the sampler to your original latent instead, TTM silently never happens, because the packed keys never reach the guider.