artiva-ai/modelos

Artiva: ONNX models for Creative Tools

ONNX conversions of open image models, used on-device (onnxruntime-web, WebGPU/WASM) by Creative Tools from Artiva.

The weights belong to their original authors. Every file keeps the license of the model it comes from; the only changes are the conversion to ONNX listed below.

File	Original model	Authors	License
scunet_color_real_psnr_512.onnx	SCUNet (scunet_color_real_psnr)	Kai Zhang et al.	Apache-2.0
fbcnn_color_512.onnx	FBCNN (color)	Jiaxi Jiang et al.	Apache-2.0
fbcnn_color_512_mixto.onnx	FBCNN (color), mixed precision	Jiaxi Jiang et al.	Apache-2.0
rasgunos_old_photos.onnx	Bringing Old Photos Back to Life, scratch detection	Ziyu Wan et al., Microsoft	MIT
gfpgan_v14.onnx	GFPGAN v1.4	Xintao Wang et al., Tencent ARC	Apache-2.0
ben2_base_web.onnx	BEN2 Base	Prama LLC	MIT
lama_512.onnx	LaMa (Big-LaMa)	Roman Suvorov et al., Samsung AI Center	Apache-2.0
lama_1024.onnx	LaMa (Big-LaMa)	Roman Suvorov et al., Samsung AI Center	Apache-2.0
isnet_general.onnx	ISNet / DIS (isnet-general-use)	Xuebin Qin et al.	Apache-2.0
birefnet_dynamic_1024_fp16.onnx	BiRefNet (BiRefNet_dynamic)	Peng Zheng et al.	MIT
birefnet_lite_1024.onnx	BiRefNet (BiRefNet_lite)	Peng Zheng et al.	MIT

Backup copies (unchanged) of third-party ONNX files the app also uses:

File	Original	License
selfie_multiclass_256x256.onnx	MediaPipe Selfie Multiclass (Google), ONNX by senty-au	Apache-2.0
depth_anything_v2_small_fp16.onnx	Depth Anything V2 Small, ONNX by onnx-community	Apache-2.0

Changes made

• All: exported from the original PyTorch weights to ONNX, checked against PyTorch on the same inputs.
• SCUNet, FBCNN: fixed 512 × 512 input (the app processes large images in tiles).
• fbcnn_color_512_mixto.onnx: convolutions in float16, quality-factor branch in float32.
• GFPGAN: the random noise inputs are fixed, so the output is deterministic.
• BEN2: Pow with a float64 exponent replaced by Mul; float64 casts and constants turned into float32, so it runs on WebGPU.
• LaMa: fixed 512 × 512 (lama_512) or 1024 × 1024 (lama_1024) input; the FFT of the Fourier units is written with MatMul/Cos/Sin (Carve-Photos/lama, FourierUnitJIT) so it runs on WebGPU.
• ISNet: exported from the official isnet-general-use.pth weights of the DIS repository with a fixed 1024 × 1024 input.
• BiRefNet: fixed 1024 × 1024 input; the deformable convolution is written with GridSample (instead of GatherND/ScatterND) so it fits in onnxruntime-web; birefnet_dynamic_1024_fp16 runs in float16.

Licenses

• Apache-2.0: https://www.apache.org/licenses/LICENSE-2.0
• MIT: see the original repositories linked above for the copyright notices.

See LICENSE for the notices.