CLIP ViT-B/32 aesthetic predictor (LiteRT)
OpenAI CLIP ViT-B/32 image encoder (QuickGELU) + LAION aesthetic predictor v1 linear head
(sa_0_4_vit_b_32_linear.pth, https://github.com/LAION-AI/aesthetic-predictor), exported as one
TFLite model with float16 weights.
- Input: NHWC float32
[1, 224, 224, 3], sRGB in[0, 1](CLIP normalization is inside the model). - Output: float32
[1, 1], aesthetic score on a roughly 1–10 scale.
Both upstream components are MIT licensed.