nx_image_feature_extract.nx

nx_image_feature_extract.nx -- fixed-length feature vector emitter. Composes existing image-quality detectors (nx_aesthetics + nx_self_similarity + nx_symmetry_group) into a 16-element Q10 feature vector. Pair with nx_cosine_similarity for image-to-image comparison, concept-embedding lookup, clustering, recommendation. FOUNDATION FOR THE "AI GROWS BY INGESTING IMAGE" ROADMAP: user uploads N images of a new outfit -> nx_image_feature_extract on each -> N feature vectors -> average them (centroid) -> "outfit_X" concept embedding -> store as content-addressed manifest in nishi-library later prompt mentions "outfit_X" -> blend stored embedding into prompt embedding (IP-adapter path) -> render conditioned on the new concept No LoRA training, no base-model fine-tune. The substrate GROWS. Idea-provenance (patent-clean): CLIP image-encoder pattern (Radford 2021) IP-Adapter (Ye 2023) Textual Inversion (Gal 2022) The architecture absorbs the IDEA -- fixed-length embedding per image -- without copying any third-party code. Today's features are substrate-native (aesthetics + fractal + multi-axis symmetry) rather than CLIP-encoder weights. VECTOR LAYOUT (each Q10): [ 0] COLOR_HARMONY hue distribution narrowness [ 1] LIGHTING exposure / range / clipping composite [ 2] COMPOSITION thirds + symmetry + balance composite [ 3] CLARITY salience peakedness [ 4] SHARPNESS mean Sobel magnitude [ 5] AESTHETIC_COMPOSITE mean of [0..4] [ 6] FRACTAL_DIM box-counting Q10 of (D-1) for D in [1, 2] [ 7] FRACTAL_AESTHETIC Spehar 2003 D~1.4 peak-fit [ 8] EDGE_DENSITY Sobel-thresh fraction [ 9] SYM_COMPOSITE max across 7 symmetry axes [10] SYM_H horizontal bilateral [11] SYM_V vertical bilateral [12] SYM_DIAG diagonal mirror