| images | IMAGE | | Source photo batch — typically 4 images: front, left, right, rear. |
| landmarks_batch | BD_LANDMARKS_BATCH | | From BD_FaceLandmarks. Provides per-view landmark positions for the similarity transform. |
| front_index | INT | -1-1–63 | Image index for the FRONT slot. -1 = auto. |
| left_index | INT | -1-1–63 | Image index for the LEFT slot (subject's left, lands at strip-right by convention). -1 = auto. |
| right_index | INT | -1-1–63 | Image index for the RIGHT slot (subject's right, lands at strip-left). -1 = auto. |
| rear_index | INT | -1-1–63 | Image index for the REAR slot. -1 = auto. |
| strip_width | INT | 4096512–16384 | — |
| strip_height | INT | 1024256–4096 | — |
| front_center_x | FLOAT | 0.500–1 | Strip-X column for the FRONT view's eye-center (0.5 = strip center). |
| left_center_x | FLOAT | 0.680–1 | Strip-X column for the LEFT view (subject's left face). 0.68 lets the left view's hull overlap the front by ~⅓ — that's the overlap region Qwen smooths and where shared landmarks meet. |
| right_center_x | FLOAT | 0.320–1 | Strip-X column for the RIGHT view. 0.32 mirrors left_center_x — same overlap with front. |
| eye_target_y | FLOAT | 0.400–1 | Strip-Y row where every view's eye-center lands. 0.4 = 40% from top. Lower → more headroom above; higher → more neck/chin below. |
| face_height_ratio | FLOAT | 0.600.1–1 | Target face height (chin↔forehead distance) as a fraction of strip height. 0.60 makes the face large enough that adjacent views overlap meaningfully at their face hulls (which is where shared landmark alignment happens). |
| align_to_frontopt | BOOLEAN | true | When ON: each side view's transform is anchored to a shared landmark (subject's-left tragion for the LEFT view, subject's-right tragion for the RIGHT view) so it lands at the SAME strip-space position as the front view placed it. This is literal 'align by face points' stitching — the left side of the front photo and the inner edge of the left-profile photo meet at the same subject feature. When OFF: each view independently targets its configured center_x (looser fit but easier to position manually). |
| feather_radiusopt | INT | 300–200 | Gaussian-blur radius for the per-view face-hull alpha. 30 = moderate feathered selection — soft enough that adjacent views crossfade smoothly in the overlap region, but each view's face is still visible as itself. Increase for more blending; decrease (e.g. 10) for visible seams. |
| rear_extentopt | FLOAT | 0.150–0.3 | Fraction of strip width per rear half (each side). |
| rear_split_xopt | FLOAT | 0.500–1 | Column in the rear photo to split for the two halves. |
| rear_flip_lropt | BOOLEAN | false | Swap which rear half goes to which strip edge. |
| rear_maskopt | MASK | | Head silhouette for rear photo (e.g. from BD_SAM3MultiPrompt('head')). When provided, each rear half is perspective-warped so its silhouette fills the slot from outer edge to back-midline. Without it, rear is plain cv2.resize-stretched. |
| rear_confidenceopt | FLOAT | 0.500–1 | Weight applied to the rear contribution in the alpha-weighted composite. |
| rear_blend_bandopt | FLOAT | 0.060–0.2 | Smooth-blend band width at the rear-slot boundary. |
