DotoPototo/paint-models

doto.gg Paint selection models

Tap-to-select models for the doto.gg Paint wall colour visualiser, which runs them in the browser with onnxruntime-web (WebGPU, with a WASM fallback). Every file is full precision (fp32): fp16 and int8 versions were tried and dropped because they made selections worse.

The app pins a commit of this repo, so a file here never changes under it: a new version means a new commit.

Folder	Model	Encoder	Decoder	Used for
mobilesam/	MobileSAM	28.2 MB	16.5 MB	Phones, tablets and slower computers
sam2.1-small/	SAM 2.1 Hiera Small	138.0 MB	16.5 MB	Desktops with a capable GPU

mobilesam

Original MobileSAM weights: Apache-2.0 (LICENSE). ONNX exports: MIT (LICENSE.MIT).

• encoder.onnx: https://huggingface.co/comicbox/mobilesam-onnx/resolve/main/mobile_sam_image_encoder.onnx, unchanged. SHA-256 580f5fb648ea1062c0aabc26217aed56921985f03f0cbbd852bba81d760cc749. Input input_image [H, W, 3] float RGB 0–255 with the longest side 1024; normalisation and padding are inside.
• decoder.onnx: https://huggingface.co/Acly/MobileSAM/resolve/main/sam_mask_decoder_multi.onnx, unchanged. SHA-256 8976b90a87ba50a6a72217a5ff994f7d25ce16f2229fcc1ed259e1294c622ffe. SAM's ONNX decoder interface, returning all four mask candidates.

sam2.1-small

SAM 2.1 (Meta): Apache-2.0 (LICENSE). ONNX export by Viet-Anh Nguyen with samexporter (MIT), from https://huggingface.co/vietanhdev/segment-anything-2.1-onnx-models/resolve/main/sam2.1_hiera_small_20260221.zip (Apache-2.0).

• encoder.onnx: SHA-256 4b71c31308ff1e49f812c10fef910f0a4011991cf4935c0a53284b79fcd1b3c3. Input image [1, 3, 1024, 1024], the photo stretched to 1024² and ImageNet-normalised. Changed from the export: its two If nodes, whose branches declare different output shapes (which onnxruntime-web rejects), are replaced by the branch a 1024² input always takes. Outputs are bit-identical to the original.
• decoder.onnx: unchanged from the export. SHA-256 d158eb26a43d39eed23ef677d6826d17eb9eda951921f9a252fa5d36d871b57d. Returns three 256² mask candidates on the stretched 1024² frame and their predicted IoU.