# Identity training and honest layers

## Soul ID — researched 2026-09-21
Current official help: https://higgsfield.ai/creator-hub/help-center/ai-models/how-do-i-create-and-use-a-soul-id-character
The help center describes 20–80 photos and explicitly allows generated portraits for invented characters. CLI help advertises 5–20 input images. Twenty is the overlap; validate the specific model route before spending. The docs prefer single subjects, varied angles/expressions, unobstructed faces and at least one full-height view. Do NOT upload an entire multi-person or multi-view editorial sheet as a training example. Crop one complete subject per image; omit captions, neighboring characters, malformed anatomy and repeated near-identical views. A crop is not independent new evidence. No claim that two crops per character here are a ready training dataset.

Soul ID is native to the Soul family; generated identity can reach other supported tools through Elements. It is not a verified drop-in identity layer for Sunburst. Trained identity is not exportable as ASH model weights. Recommendation: no training is needed to ship v0.1. Evaluate one curated character only when repeated-shot consistency is the actual bottleneck, with separately approved cost. No training or upload was performed for this pack.

## Transparency
Six single-character cutouts were extracted locally from approved reference sheets. Masks were inspected against light and dark backgrounds. Auri uses a higher-quality local extractor to retain opaque pale tail fur; Sable's actual tail loop was restored from the source mask. Edge pixels remain an approximation: no invisible anatomy or occluded costume was reconstructed. Source crops and alpha outputs are retained separately.

## Layers
OpenRaster (.ora) files open in compatible editors such as Krita and GIMP. Character ORAs contain an extracted subject, a newly made neutral backdrop, and a hidden original flattened crop. Location ORAs split the original page into establishing and supporting-image regions. These are EDITABLE PRESENTATION LAYERS, not original brush strokes, vector costumes, 3D depth or recovered hidden scenery. Native flattened originals remain untouched. Fonts/labels in generated raster artwork remain pixels.