MimicMotion Decode

If MimicMotion Sampler is the engine, this is the tailpipe. MimicMotion Decode takes the LATENT output of the Sampler and runs it through the pipeline's VAE to produce an actual batch of images - the thing you can feed to a Save Video node, inspect, upscale, or face-fix. It's the least glamorous node in the pack and also the one you literally cannot finish a workflow without.

The mechanism is boring in the best way: decode the latent frames through the SVD temporal decoder, convert back to standard ComfyUI image tensors. Two details are worth knowing. First, it deliberately drops the first frame of the decoded output - the reference frame, which MimicMotion's pipeline repeats as frame zero. Your output starts at the first driven frame, which is what you want, but it catches people off guard when their clip looks one frame short. Second, decoding happens in chunks to keep VRAM under control.

Inputs and output

• mimic_pipeline - the MIMICPIPE handle from (Down)Load MimicMotionModel. Same object the Sampler used; they share it.
• samples - the LATENT from MimicMotion Sampler.
• decode_chunk_size - how many frames the VAE decodes at once. Default 4. If you hit out-of-memory during decode (and you will, on some GPUs), drop this to 2 or 1. If the chunked decode fails outright, the node silently retries with chunk size 1, which is a nice touch from Kijai.

Output is a single IMAGE tensor (images) - a batch of frames in order, ready for a SaveVideo-style node or whatever post-processing you have lined up.

Install

The whole pack comes together; there's no separate install for this node. ComfyUI Manager → search "ComfyUI-MimicMotionWrapper" → install, or:

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

That pulls diffusers, transformers, and accelerate. The actual model downloads (the ~3GB MimicMotion UNet and ~4GB SVD package) happen on the loader node's first run, not here.

Gotchas

Not much to it, honestly. The decode is fast relative to sampling, so the only failure mode worth remembering is VRAM - and that has a one-knob fix (decode_chunk_size). The other thing worth knowing is that this node's output is raw 2024-era MimicMotion quality: smooth motion, character consistency you can rely on, but mushy faces and soft detail. That's the model, not the decoder. Most people who finished these workflows did a face pass or a second refinement pass on exactly these frames.