(Down)Load MimicMotionModel

Every MimicMotion workflow in this pack starts here. (Down)Load MimicMotionModel is the loader node, and the "(Down)" in its display name is doing real work: unlike a plain loader, this thing will happily download several gigabytes of models on first use and stash them where ComfyUI expects them. It's Kijai's wrapper around Tencent's MimicMotion, a 2024 pose-controlled image-to-video model built on Stable Video Diffusion 1.1 - so don't expect this to be bleeding edge. It's a well-made museum piece.

What it actually loads

One node call assembles the entire inference pipeline:

• The pruned MimicMotion UNet from Kijai/MimicMotion_pruned on HuggingFace, downloaded to ComfyUI/models/mimicmotion. This is the SVD UNet with MimicMotion's pose conditioning baked in.
• The SVD XT 1.1 diffusers package (VAE, image encoder, scheduler, feature extractor) into ComfyUI/models/diffusers/stable-video-diffusion-img2vid-xt-1-1. Fun detail: the code deliberately ignores the unet files from that repo, because the MimicMotion pruned file is the UNet. You're downloading ~4GB of supporting cast, not a second UNet.
• The PoseNet weights from the pack's own models folder.

Everything comes back as a single MIMICPIPE output - one opaque handle you feed to both MimicMotion Sampler and MimicMotion Decode. You never touch the pieces directly.

The two inputs

• model - MimicMotionMergedUnet_1-0-fp16.safetensors vs the _1-1_ variant. The 1.0 model plays nice with the default 16-frame context; the 1.1 model expects long clips - the Sampler will warn you to set context_size to 72 if you load it and don't. For casual use, 1.0 is the one you reach for.
• precision - fp16 is the default and the right call almost every time; the pruned weights ship as fp16. fp32 upcasts and eats VRAM for no visible gain. bf16 is there for experimentation on cards that support it. If you're not sure, leave it on fp16.

First-run gotchas

The download happens the moment the node executes, and it's big - around 3GB for the UNet plus ~4GB for SVD, on top of the DWPose models the pose node will pull separately. That first run will look hung for a while. It isn't; it's downloading. Watch the ComfyUI console for the Downloading model to: ... lines.

The flip side: because models land in models/mimicmotion and models/diffusers, they persist across restarts and are reused. The node only re-downloads if the files are missing.

Install

Standard Kijai pack, two routes. ComfyUI Manager - search "ComfyUI-MimicMotionWrapper" and install. Or manually:

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

On the Windows portable build, install into the embedded Python: python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\ComfyUI-MimicMotionWrapper\requirements.txt. The dependencies themselves are light - diffusers>=0.28.0, transformers, accelerate - so the heavy lifting is storage and bandwidth, not pip. Note the README says "WORK IN PROGRESS" and the repo's last commit was January 2025; this pack is finished, in the sense of "done being developed."

One honest caveat to end on: you'd only set this up today if you specifically want the MimicMotion look. For current pose-driven video work the ecosystem has moved to Wan/VACE and LTX-2, and this loader will feel like a time capsule. A fun time capsule, but a time capsule.