aether-models/aether-kit-clip-vit-b32

CLIP ViT-B/32 for Aether-Kit (Core AI)

openai/clip-vit-base-patch32 at revision 3d74acf9, exported by Aether-Kit's Forge exporter (Forge/clip/) as two static Core AI graphs: the image tower (an RGB image resized to 224 on its shortest side and center-cropped, on [0, 1]; CLIP normalization is in the graph) and the text tower (77 CLIP BPE tokens, batches of 1 or 8). Each returns unit-length 512-wide embeddings in float32; the towers run in float16. Cosine similarity is the dot product, and zero-shot classification is a softmax over 100 x cosine. Load it with Aether-Kit's AetherVision (clip-joint@1, macOS); the manifest is .aether/aether.json. Gated against the upstream implementation on CPU in float32 (see export-report.json).

OpenAI's model card describes CLIP as a research model and considers untested deployed use, surveillance and facial recognition out of scope.

MIT (LICENSE). A modified version (see NOTICE).