Motion Bridge Save ????

WAN 2.2 tops out around 81 frames of native context, so any longer video is segments bolted together - and the bolt is the problem. If the next segment starts from a frozen frame with no motion memory, you get a visible jump every single time. IAMCCS_MotionBridgeSave writes the state that fixes that: it takes the tail of your finished segment's latents and the last frame, and saves them to disk as a "motion bridge" your next workflow can load.

It's the save-side of the pair with IAMCCS_MotionBridgeLoad, and it's an output node - meaning it ends a branch, saves, and passes nothing onward. This is deliberate: it's meant to sit at the end of segment N's workflow, producing the hand-off file for segment N+1.

How it works

Give it the latent from your segment's decode/sample path and the matching images, and it does three things:

1. Slices tail_frames latent frames off the end of latent - those carry the motion state the next segment needs as its motion reference.
2. Saves everything as output/motion_bridges/<name>.safetensors.
3. Optionally saves <name>_lastframe.png, the junction frame, for visual inspection and as a fallback LoadImage source.

The flush_vram toggle (default on) moves the input tensors to CPU and calls ComfyUI's soft empty cache plus torch.cuda.empty_cache() after saving - a real courtesy on tight cards, where one more KSampler OOM right after a save is the classic way a long run dies. Turn it off only if a GPU branch runs immediately after this node and needs the tensors resident.

Inputs and outputs

• latent (LATENT) - required, the segment's latents.
• images (IMAGE) - required, the decoded frames.
• name (STRING, default segment_bridge) - filename without extension. The tooltip suggests descriptive names like scene01_seg2_to_seg3; follow it, because six bridges named segment_bridge is how you lose an afternoon.
• tail_frames (INT, default 2) - how many latent frames to slice. Must match motion_latent_count on the next workflow's WanImageMotion-style node.
• save_preview_png (BOOLEAN, default true) - save the last frame as a PNG alongside.
• flush_vram (BOOLEAN, default true) - free CUDA memory after saving.

Outputs: none. It's a terminal node.

How to install

Ships in IAMCCS-nodes. ComfyUI Manager search "IAMCCS", or:

cd ComfyUI/custom_nodes
git clone https://github.com/IAMCCS/IAMCCS-nodes.git

Restart. No model downloads. Pack floors: ComfyUI ≥ 0.3.0, Python ≥ 3.12, Torch ≥ 2.8.

Gotchas

The contract that matters is tail_frames: the number you slice here has to equal the motion_latent_count your next segment's motion node expects, or the continuity logic reads the wrong amount of tail and the seam drifts. If you change one, change the other. And remember the load-side tooltip's warning applies in reverse - newly saved bridges don't appear in the load dropdown until the ComfyUI page is refreshed. Also, this saves only motion/latent state, not character identity; for long-form where faces drift across segments, that's still on you (the Wan community's usual answer is a character LoRA plus the FF/LF anchoring this pack made popular).