This demo uses the BioCLIP backbone (imageomics/bioclip via open_clip) with a frozen encoder and a small linear head trained on the BioTrove Reptilia subset (189 species). Inputs are normalized with the BioCLIP validation transform, fed through the encoder, and scored with a softmax over the reptile classes.

How inference works here:

• Uploaded images are converted to RGB, preprocessed with the BioCLIP pipeline, and run on the available device (CPU/GPU).
• Live camera frames are downscaled (~640px wide), captured every ~1.5s, and sent to /predict_frame for scoring.
• Top-k predictions come from the softmax probabilities over the 189-class head.

Only the BioCLIP model is used for both uploads and live feed; no additional detectors or models are involved.