Animate image with Animate_X_v2

Animate image with Animate_X_v2 is the upgraded Animate-X img2vid node, and it fixes two things at once. First, it adds the Implicit Pose Indicator (IPI) from the Animate-X paper - an extra pose embedding alongside the Explicit Pose Indicator - which the author reports gives a "slight improvement" in output quality over the original node. Second, it folds pose generation in: like the Long variants, v2 extracts its own DWPose from the video input internally, so you no longer need to wire in a Gen_align_pose node ahead of it.

That structural change is the bigger deal for beginners. The base Animate-X node (Animate_X_Image) demands three separate inputs from the pose-prep node; v2 takes just image and video and does the rest. It also changed which settings it trusts by default - useFirstFrame and dontAlignPose both default to true here, versus false on the v1 nodes. That's the author choosing the more robust path for you: treat the driving video's first frame as the reference, and skip pose alignment on the assumption that you'll frame things yourself.

Inputs and outputs

• image - your reference photo of the person.
• video - the driving clip; its own pose sequence is extracted internally.
• useFirstFrame (default true) - use the video's first frame as the identity reference instead of image.
• dontAlignPose (default true) - leave true if the people already match in size; flip false to auto-scale the driving poses to your character.
• frame_interval (1–8, default 1), max_frames (default 32), seed, steps (default 25 - slightly fewer than v1's 30), resolution_x (512 or 768).

Outputs are video (the frame batch) and poses (the internally generated DWPose sequence - the Animate-X version, so you can sanity-check what the model was conditioned on).

Setup

Same single-extra-file story as v1: install the pack (Manager search "UniAnimate Nodes for ComfyUI" or git clone https://github.com/Isi-dev/ComfyUI-UniAnimate-W into custom_nodes/), make sure the ~14GB UniAnimate stack is in checkpoints/, and drop in animate-x_ckpt.pth - the only file specific to the Animate-X nodes. Python 3.9+, xformers, and onnxruntime are prerequisites.

Which Animate-X node should you run?

If you're starting fresh, v2 is the sensible default: it's the newest, adds the paper's IPI conditioning, and removes a node from your graph. The v1 node (Animate_X_Image) still works and is what you'll find in older shared workflows - it takes explicit ref_pose/pose_sequence inputs from Gen_align_pose, which is genuinely useful if you want to reuse a precomputed pose sequence across several generations. The VRAM profile is the same ~12GB ballpark (trim max_frames on smaller cards), and expect roughly eight minutes per 32-frame batch at [512, 768]. v2 is the pick; v1 remains the fallback for workflows you didn't build.