ComfyUI Face-Aligned Center

Batch-aware character centering for ComfyUI. Point it at a character sheet (typically 8 angles: front / 3-4 / side / back / etc.) and every image comes out with the face at the same size and the same canvas position — including back-view angles that don't have a visible face.

How it works

1. Auto-detect BG color from the image border (works for white, green, black, any uniform fill), then build a foreground mask. Max-channel distance threshold — same behavior as the classic min(RGB) < 245 rule on white BG, generalized to any color.
2. Silhouette analysis per image: trim thin weapon-tip rows off the top and bottom, then find the head x-center using per-row longest-contiguous-run midpoints (robust against sceptres/swords that extend sideways).
3. One batch scale for the whole batch: target_body_h / median(body_h). Every image uses the same scale so all angles end up at the same physical size — including ones with extended props that made their own silhouette bigger.
4. Translate so the silhouette head-cx lands at the user-specified canvas position.
5. Padding on the output canvas is filled with the same BG color that was detected on the input, so a green-screen image comes out on a clean green canvas, a white-BG image stays white, etc.

No face-detection model required — the silhouette head-cx is stable even on back views where a face detector would fail.

Nodes

Face-Aligned Center

Inputs:

• images (IMAGE batch) — the character sheet frames
• canvas_size (INT, default 928) — square output size in pixels
• face_fill (FLOAT, default 0.12) — face height as a fraction of the canvas. The one zoom knob. Lower = more empty space around the character, higher = tighter crop.
• face_y (FLOAT, default 0.30) — vertical position of the face center on the canvas (0 = top edge, 1 = bottom edge).
• mask (MASK, optional) — pre-computed foreground mask. Overrides the BG-color auto detection.

Outputs:

• images (IMAGE batch) — centered/scaled results, same count as input.
• masks (MASK batch) — the foreground mask after scaling/translation.
• info (STRING) — diagnostic: batch scale + per-image detected head coordinates.

Anime Face Detect & Crop

Detect an anime/chibi face using lbpcascade_animeface (OpenCV Haar cascade, ~2MB, CPU-only, auto-downloaded on first use) and output a cropped face image plus a rectangle mask at the original image dimensions. Useful for ID/reference extraction and inpainting masks.

Inputs:

• images (IMAGE batch)
• crop_size (INT, default 512) — output square crop size.
• padding (FLOAT, default 0.3) — extra margin around the bbox.

Outputs:

• cropped_images (IMAGE batch) — square face crop at crop_size. Falls back to a center crop if detection fails.
• face_masks (MASK batch) — rectangle at the original image dimensions.

Face-Aligned Fine-Tune

Manual per-image adjustment on top of the auto-aligned output. Interactive canvas preview with zoom / nudge buttons — all changes render client-side (no workflow queue triggered per click), so iteration is instant. A toggleable grid overlay (crosshair + diagonals + safe-box) helps verify alignment. The overlay is view-only and never reaches the clean IMAGE output.

Inputs:

• images (IMAGE batch) — typically from Face-Aligned Center.
• zoom / dx / dy — driven by the + − and ← ↑ ↓ → buttons. Auto-reset when a new batch of images flows in, so a pipeline running many times never applies stale adjustments to a fresh character.

Outputs:

• images (IMAGE batch) — zoom + translate applied, no grid baked in.

Installation

Via ComfyUI Manager, or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/serhiiyashyn-sf/comfyui-face-aligned-center
cd comfyui-face-aligned-center
pip install -r requirements.txt

License

MIT