Jingzao Image Forge v1.10.0
Reference distillation now retains selected hair, makeup, expression, gaze, pose, action, body, wardrobe and photography instead of reducing the reference to color and material. The optional --reference-profile path separates observed evidence, target choices and positive generation prose, with a facet-level audit kept outside the prompt.
- Incorporates DIR's evidence-to-controls/review method without requiring or modifying the upstream DIR executor.
- Includes four source-profile branches and text-only refinement examples. Unknown source regions cannot be projected as preserved or adapted facts; target-only original design stays in the base specification.
- Preserves existing specs and style-capsule behavior; profile and capsule flags are mutually exclusive.
- Rebuilds the English/Chinese README landing pages, guide, AI-readable llms.txt map and repository discovery metadata, with newly generated introduction artwork and a current public sample. Older graphics remain clearly historical.
Validation: 251 tests, the full local workflow, PR/main/tag CI, independent code and visual review, live README checks, and a fresh-process installed-compiler smoke test. All 164 tracked files and executable modes match the merged release commit and ZIP.
Four female profile-generation calls returned images. Attached-reference trials retained their main traits; the text-only refinement improved finger placement and ribbons but still varied in fringe/material and cropped the hat top. A separate window-side body/hand trial was blocked at input moderation and returned no image. These are different inputs from the v1.9 failures: no moderation root-cause fix, universal acceptance, full-body validation or independent identity guarantee is claimed. The older generic capsule remains draft.
The ZIP contains the jingzao-image-forge/ Skill folder. User source images and source-attached portrait outputs are not distributed. One original text-only sample and the two new introduction layouts are included.