asciiartme-models

ONNX models used by asciiart.me (aame), which turns pictures into Japanese ASCII art (AA). The browser version loads them from here; the Python version can also get the same files from the repository's release models-v2 (python -m aame.get_models).

They were converted with python -m aame.export_onnx from the original weights, without retraining:

File Weights Licence
manga_line.onnx erika.pth from ljsabc/MangaLineExtraction_PyTorch (release v1), with its weights stored as float16 (computed in float32; 86 MB instead of 173 MB) MIT
teed.onnx checkpoints/BIPED/7/7_model.pth from xavysp/TEED MIT
face_unet.onnx model/UNet.pth from siyeong0/Anime-Face-Segmentation MIT
isnet_anime.onnx isnetis.onnx from skytnt/anime-seg (SkyTNT/anime-segmentation), with its weights stored as float16 (computed in float32) and ceil_mode turned off in its MaxPool nodes (no effect at 1024x1024; WebGPU needs it) Apache-2.0

THIRD_PARTY_NOTICES.txt holds these licences; keep it with the model files.

Inputs and outputs

For the first three, H and W are free, but must be multiples of 64. The pre- and post-processing is in aame/models.py.

File Input Output
manga_line.onnx gray 1x1xHxW, grey 0..255 line 1x1xHxW, black lines on white (may go outside 0..255)
teed.onnx rgb 1x3xHxW, RGB 0..255 edge1, edge2, edge3, edge_fused, each 1x1xHxW, before the sigmoid
face_unet.onnx rgb 1x3xHxW, RGB 0..1 parts 1x7xHxW, probabilities
isnet_anime.onnx img 1x3x1024x1024, RGB / max βˆ’ (0.485, 0.456, 0.406) mask 1x1x1024x1024, foreground (as rembg's isnet-anime)

sha256

d486432cada895821b1153a92a64f2ffe9f21bf97e46a8c1c1bdb301de1da20f  manga_line.onnx
0ad443441475c06192309b0649b86ab6e6c1cc88cc6441e08ce7a6afa2c3e78c  teed.onnx
acc4b2c0a246a93bb643c8f626ebf4042e1113657fcc22d6286a79f74e656efe  face_unet.onnx
fdf90b7af006d665c6db2f657b5ef0a2f6e13dab010acb6a223fbc18d04db068  isnet_anime.onnx
0f9e94847aab0addfd89c34ba9c0f5fd3429d594f99e998503f7129ce388cd52  THIRD_PARTY_NOTICES.txt
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support