Tossana ទស្សនៈ · Khmer image captioning

Upload a photo and Tossana describes it in Khmer. 104M parameters: a SigLIP image encoder and a small Khmer decoder. Try both versions: the ONNX int8 file that runs on a phone (102 MB), or the full PyTorch model (the same model trained on GPU, running here on CPU). Preview: it can name objects that are not in the photo.
Model · ONNX · Website

Photo

Model

ONNX int8 · small & fast (102 MB, on-device) PyTorch · full model (417 MB)

Best accuracy (beam search, a bit slower)

Tossana writes

Example photos (Wikimedia Commons, open licenses)

Photo	Model	Best accuracy (beam search, a bit slower)

Example photos: Pwnsey, CC BY-SA 3.0 · Dickelbers, CC BY-SA 3.0 · Adam Jones, CC BY-SA 2.0 · Anandajoti, CC BY-SA 3.0 · Hiart, CC0 · Jakub Hałun, CC BY-SA 4.0 · Marcin Konsek, CC BY-SA 4.0 · Rik Schuiling / TropCrop-TCS, CC BY-SA 4.0