Image Resize WAN Adaptive Framing

Image-to-video with Wan (2.1/2.2, the local video workhorse) has two silent killers: a start frame at the wrong resolution, and a subject framed so badly that the model invents or mangles it the moment motion starts. Image Resize WAN Adaptive Framing handles both in one pass. It auto-selects a Wan-friendly resolution, detects faces and bodies, and applies subject-aware framing via crop or soft-bar padding before you ever hit the video model. You drop a source image in, get a properly prepped start frame out, plus the exact width/height and the preset it chose.

This is a "prepare, don't guess" node for the Wan I2V pipeline. The pack also ships Wan LoRA-by-text loaders for the same workflow, so this is clearly the author's own image-to-video prep ritual, packaged.

How it works

Three stages, each grounded in real choices:

Resolution selection. Wan presets in three tiers - small (416×240), main (832×480), high (1280×720) - with vertical variants derived by transposing. resolution_tier defaults to auto, which picks the tier closest to your source's area, and orientation_mode (auto/force_horizontal/force_vertical) decides the orientation. This keeps you on the resolutions Wan was trained on instead of random multiples.

Subject detection. The clever bit: it uses OpenCV Haar (faces) + HOG (people) detectors - lightweight CPU methods, deliberately not a deep model, per the source notes - so no VRAM and no checkpoint download. framing_mode defaults to auto_subject: face first, body if no face, center if neither. The detection dials (face_min_percent, detect_scale_factor, detect_min_neighbors) match the pack's face-crop node.

Framing. anchor_strength (0.85) and top_bias (0.18) position the anchor, then resize_mode applies it as crop (content-preserving crop to the preset) or pad (fit within the preset with soft bars). upscale_method (default lanczos) picks the interpolator.

Outputs: resized_image, debug_image (an overlay showing detected faces/bodies and the crop box - default on, disable with emit_debug_image), plus width, height, and selected_preset strings/ints so downstream nodes know exactly what you're working with. Those metadata outputs are the difference between this and a blind resize - wire them to a text node or your filename to see the decision.

Installing it

Part of CorvaeOboro's ComfyUI_illumorae (CC0 - no model downloads; the Haar/HOG detectors ship inside OpenCV). ComfyUI Manager → search illumorae, or:

cd ComfyUI/custom_nodes
git clone https://github.com/CorvaeOboro/ComfyUI_illumorae

Common issues

• Subjects ignored / face not found - Haar wants frontal faces. In auto_subject it falls back to body (HOG), then center; if a subject still gets cropped, check debug_image to see what was detected, and lower face_min_percent/min_neighbors.
• Wanted horizontal, got vertical (or vice versa) - orientation_mode is auto by default, so it follows the source. If your pipeline needs a fixed orientation regardless of input, force it.
• pad bars look wrong - soft bars are matte-colored padding by design; if you want true letterboxing with a specific color, the pack's Image Resize w Matte gives you explicit matte control instead.
• Video looks off anyway - the node preps the frame, but Wan 2.2 motion/quality still depends on your sampler and LoRAs. Treat this as removing a failure mode, not guaranteeing results.