Useotools models
Copies of the open models that useotools.com runs in the visitor's browser with ONNX Runtime Web. Each file sits in a folder with its own licence. Nothing here is trained by us. The only change is the ISNet file, whose weights are stored in 7 bits per channel to make the download smaller.
| Folder | File | Model | Licence | Original |
|---|---|---|---|---|
isnet/ |
isnet-general-use-w7.onnx.gz |
ISNet general-use, background removal (weights stored in 7 bits, gzipped) | Apache-2.0 | xuebinqin/DIS via rembg |
migan/ |
migan_pipeline_v2.onnx.gz |
MI-GAN inpainting pipeline (gzipped, otherwise unchanged) | MIT | Picsart-AI-Research/MI-GAN, andraniksargsyan/migan |
lama/ |
lama_fp32.onnx |
big-LaMa inpainting, 512 x 512 | Apache-2.0 | advimman/lama, Carve/LaMa-ONNX |
mobilesam/ |
mobile_sam_image_encoder.onnx, sam_mask_decoder_single.onnx |
MobileSAM encoder and Segment Anything mask decoder | Apache-2.0 | ChaoningZhang/MobileSAM, Acly/MobileSAM |
upscale/ |
realesr-general-x4v3-dn0/dn50/dn100.onnx, realesrgan-x4plus.onnx, realesrgan-x4plus-anime-6b.onnx |
Real-ESRGAN 4x upscaling: general (three noise levels), x4plus, anime 6B | BSD-3-Clause | xinntao/Real-ESRGAN, warpstead/enhance-models, deepghs/imgutils-models |
colorize/ |
ddcolor-tiny-fp16w.onnx |
DDColor-tiny, black and white photo colourisation, 512 × 512 (edgetools' fp16 export, MODEL-CARD.md, with each fp16 weight cast to float at load by useotools' scripts/onnx-fp16-storage.py: fp16 maths gave wrong colours on WebGPU) | Apache-2.0 | piddnad/DDColor, ONNX export edgetools/ddcolor |
whisper/base/, whisper/small/, whisper/large-v3-turbo/ |
ONNX graphs under onnx/ plus their config and tokenizer files, in the layout of the onnx-community repos |
OpenAI Whisper speech recognition (audio to text): base and small for the processor and the graphics card, large-v3 turbo (q4f16) for the graphics card | base, small: Apache-2.0; large-v3 turbo: MIT | openai/whisper, ONNX by onnx-community (revisions 1846881b, 36050c46, 360ebcde) |
translate/<pair>/ (18 pairs: en-hi, hi-en, en-ar, ar-en, en-zh, zh-en, en-es, es-en, en-fr, fr-en, en-de, de-en, ru-en, ja-en, ko-en, id-en, mul-en, en-mul) |
onnx/encoder_model_quantized.onnx, onnx/decoder_model_merged_quantized.onnx plus config and tokenizer files, in the layout of the Xenova repos |
Opus-MT text translation (translate image), int8, about 110 MB a pair | CC-BY-4.0 or Apache-2.0 (each model's card; see src/lib/translate/files.js) | Helsinki-NLP/Opus-MT, ONNX by Xenova (revisions in src/lib/translate/files.js; fetched by scripts/translate-models.js) |
All copyright stays with the original authors. See the LICENSE file in each folder.