Mooshie/kaloscope-web

Kaloscope web (ONNX for the browser)

Browser-sized ONNX copies of Kaloscope, used by the Kaloscope 3.0 Compare Space with onnxruntime-web.

File	Source	Size
kaloscope3_v1_artist_web.onnx	heathcliff01/Kaloscope3.0-preview v1-artist-classifier/model.safetensors	169 MB
kaloscope2_web.onnx	DraconicDragon/Kaloscope-onnx v2.0/kaloscope_2-0.onnx (from heathcliff01/Kaloscope2.0)	294 MB
labels_v3.json, labels_v2.json	the matching class_mapping.csv files, as JSON arrays indexed by class id
probe_ref.json	reference outputs for a synthetic input, used by the Space to check a backend gives correct results

Kaloscope 3.0 model

Input pixel_values float32 [N, 3, 512, 512]: RGB, short side resized to 512, center crop 512, ImageNet mean/std. Outputs logits [N, 44129] (artist classifier) and style [N, 256] (L2-normalized style embedding). The classifier input normalization from the model card (L2 normalize the pooled CLS + mean patch features, times sqrt(1536)) is built into the graph.

Kaloscope 2.0 model

Input input float32 [N, 3, 448, 448]: RGB, short side resized to 512 (bicubic), center crop 448, ImageNet mean/std. Output output [N, 39261].

How these were made

export_kaloscope3.py exports the 3.0 model with torch.onnx.export (max logit difference vs PyTorch about 5e-5). shrink.py then turns Gemm layers into MatMul, quantizes every MatMul weight to 8-bit (symmetric, block size 32, weight-only MatMulNBits) and stores the remaining large conv weights as fp16. Activations stay fp32. On test images the top-1 artist matched the full-precision models, and 3.0 style vectors had cosine ≥ 0.999 to the originals. Plain int8 dynamic quantization was tried first and broke the outputs.

License

Kaloscope 3.0 is a DINOv3 fine-tune and is distributed under the DINOv3 License (LICENSE.md). Kaloscope 2.0 is Apache-2.0.