Draw ViT Pose (V2)

Wan Animate is driven by a stick figure. The model's pose conditioning channel is a rendered skeleton image, and this node is what renders it. DrawViTPoseV2 takes the pose_data from PoseAndFaceDetectionV2 and draws the whole-body skeleton - bones, hands, face, even the iris and gaze arrows if you want them - onto a clean black canvas at your target latent resolution. Whatever you see in its pose_images output is, pixel for pixel, the pose the sampler is going to animate against. So get that skeleton right.

What you set

The required inputs are almost all self-explanatory, and two of them are the ones that matter:

• width / height - render canvas in pixels, and they should match your sampler latent size (832×480 default). The pose is retargeted onto this canvas, so if it doesn't match, you get a skeleton that's the wrong size for the generation.
• retarget_padding (16) - padding around the body bbox when fitting. Larger = more headroom so big motions don't clip the frame edges. Bump it to 32–64 if limbs get cut off.
• pose_draw_threshold (0.3) - per-keypoint confidence for drawing. This is separate from the detection node's pose_threshold: detection stores everything, this decides what gets rendered. Lower to 0.1 if limbs keep disappearing from the output.
• body_stick_width / hand_stick_width - both default to -1 (auto from canvas size). 0 on hands means don't draw them at all, which is handy if hand tracking is noisy.
• draw_head - whether the face/head skeleton renders.

The genuinely interesting stuff lives in the optional inputs. face_images (wire in face_images_512 from the detection node) makes this node a one-stop shop: it validates frame-count parity, force-resizes to 512×512, and forwards the crop on its face_video output, so a single DrawViTPoseV2 can feed both the pose and face inputs of your Wan-Animate sampler. That's the wiring you actually want for a clean graph.

The outputs beyond the skeleton

pose_images is the main event. But there are two things here that no other pose-render node does:

face_video + face_cfg_scale - passthroughs. face_cfg_scale is connection-only, deliberately: it used to be a second editable widget that could silently diverge from the detection node's value, so the author made this one read-only. Unconnected it's 1.0, a no-op. And as the detection node's tooltip hammers home, Kijai's wrapper has no face-CFG input today - use WanVideoAnimateEmbeds.face_strength for real expression control.

The closed-loop critic. Wire reference_expression_coeffs_json from a source-video run of the detection node (with export_expression_coeffs=True) and this node will measure ARKit-52 blendshapes from its own pose_data - i.e. the generated side - and report per-AU error against the source. You get critic_report_json, worst_aus_csv and overall_mae (0.0 = perfect). It's a numeric fidelity signal instead of eyeballing frames, and it's zero cost when left unwired. It's overkill for most people, but if you're chasing a specific expression that keeps getting lost, it tells you which action unit to blame.

One knob that quietly matters

apply_pose_edits_to_face (default "warp") is the delivery mechanism for expression edits made upstream (Wan Face Controller 3D's expression dials, dragged landmarks, or the detection node's force_eyes_open). The reason it exists: Wan Animate's face conditioning is 100% pixel-driven - landmarks only place the crop, the LIA motion encoder reads raw crop pixels. So editing landmarks alone changes nothing visible; this node warps the actual face-crop pixels to match the edited landmarks. If you set force_eyes_open upstream and see no effect, check this is still on warp - with "off", your edits silently do nothing.

Install-wise it's just part of the pack: ComfyUI Manager (search "WanAnimatePreprocessV2") or git clone https://github.com/Code2Collapse/ComfyUI-WanAnimatePreprocessV2 into custom_nodes, then pip install -r requirements.txt and restart. No extra model downloads for this node itself - the heavy lifting happens in the detection node upstream.