Model files for Rawtonic's optional local AI tools and the ONNX Runtime libraries its opt-in build bundles. Core editing in Rawtonic needs none of them.
Every file is either byte-identical to its authors' release or exported from
their published checkpoint, by scripts/models/build_model_pack.py in the
Rawtonic repository (which checks the SHA-256 of every source). The files are
attached to the releases of this repository; models.json lists each file
with its size, SHA-256, source and license, and the app verifies the SHA-256
of every download.
| File | Used for | Source | License |
|---|---|---|---|
sam2.1_tiny_encoder.onnx, sam2.1_tiny_decoder.onnx |
Click and box masks | Exported from Meta's SAM 2.1 Hiera-T checkpoint (sam2.1_hiera_tiny.pt, facebookresearch/sam2 at 2b90b9f) by scripts/models/export_sam2.py, fp16 storage |
Apache-2.0 |
birefnet_subject.onnx |
Subject and background masks | BiRefNet general checkpoint (Peng Zheng et al., ZhengPeng7/BiRefNet, revision e2bf8e4), exported by scripts/models/export_birefnet.py, fp16 storage |
MIT |
birefnet_fp16_coreml_* (Core ML package, three files) |
Subject and background masks on macOS (builds from 2026-10-11) | BiRefNet general checkpoint (revision e2bf8e4), converted by scripts/models/convert_birefnet_coreml.py, fp16 (equal to fp32 against subject references) |
MIT |
birefnet_coreml_* (Core ML package, three files) |
Subject and background masks on macOS (builds before 2026-10-11) | BiRefNet general checkpoint (revision e2bf8e4), converted by scripts/models/convert_birefnet_coreml.py, fp32 |
MIT |
sam_vit_b_01ec64_encoder.onnx, sam_vit_b_01ec64_decoder.onnx |
Click and box masks (builds before 2026-10-09) | Exported from Meta's Segment Anything ViT-B checkpoint (sam_vit_b_01ec64.pth), per-channel uint8 weights |
Apache-2.0 |
u2net.onnx |
Foreground mask | U²-Net (Qin et al.), rembg's ONNX export, unchanged | Apache-2.0 (weights), MIT (export) |
relight_geometry_da3_base.onnx |
Relight relative scene geometry and camera | DA3 Base (ByteDance Seed, depth-anything/DA3-BASE revision f4a6c9b3c95e41c82048423d3493a81ec3fa810e), exported by scripts/models/export_relight.py; cached analysis, not metric ground truth |
Apache-2.0 |
depth_anything_v2_vits.onnx |
Depth mask | Depth Anything V2 Small, fabio-sim's ONNX export, unchanged | Apache-2.0 |
siglip2_vision.onnx, siglip2_tags.bin (and siglip2_text.onnx, siglip2_tokenizer.json for own tags) |
Automatic tags | Exported from Google's SigLIP 2 B/16 (google/siglip2-base-patch16-224, revision 75de2d5); siglip2_tags.bin holds the text embeddings of Rawtonic's built-in tag list |
Apache-2.0 |
nlmf_coreml_* (Core ML package, three files), nlmf.onnx |
AI Denoise | Nonlocal MatchFilter (Sanchez-Beeckman and Buades), Core ML and ONNX conversions by sheldonxxxx/RapidRAW-Nonlocal-Denoise (revision af6bf36), unchanged | MIT |
migan_pipeline_v2.onnx |
Object removal | MI-GAN 512 Places2 (Picsart AI Research), its first author's ONNX pipeline (revision 406830d), unchanged | MIT (weights) |
skyseg_upernet_convnext_t.onnx |
Sky mask | UperNet + ConvNeXt-T on ADE20K (Meta's checkpoint via openmmlab/upernet-convnext-tiny, revision 876ffc5), exported by scripts/models/export_sky.py, fp16 storage |
MIT |
skyseg_coreml_* (Core ML package, three files) |
Sky mask on macOS | UperNet + ConvNeXt-T (revision 876ffc5), converted by scripts/models/convert_sky_coreml.py, fp32 |
MIT |
face_detection_yunet_2023mar.onnx |
People masks (faces) | YuNet 2023mar (Shiqi Yu et al.), OpenCV Zoo at f12e127, unchanged | MIT |
face_mesh_v2.onnx, selfie_multiclass_256.onnx |
People masks (landmarks; skin, hair, clothes) | MediaPipe Face Mesh V2 and multiclass selfie segmentation (Google, version 1), converted to ONNX by scripts/models/export_faces.py |
Apache-2.0 |
libonnxruntime-*.so/.dylib, onnxruntime-*.dll |
Runtime | Microsoft ONNX Runtime 1.22.0 release archives, unchanged | MIT + third-party notices |
libonnxruntime-1.30.0-*.so, onnxruntime-1.30.0-*.dll |
Runtime on Windows and Linux (GPU plugins need ≥ 1.24.4) | Microsoft ONNX Runtime 1.30.0 release archives, unchanged | MIT + third-party notices |
*onnxruntime_providers_webgpu-0.4.0-* |
GPU for the AI models on Windows (Direct3D 12) and Linux (Vulkan) | Microsoft's WebGPU plugin execution provider 0.4.0 (PyPI onnxruntime-ep-webgpu), library unchanged; the DirectX Shader Compiler DLLs are not included | MIT (includes Dawn, BSD-3-Clause) |
*onnxruntime_providers_cuda-0.2.0-cuda13-* |
NVIDIA GPUs, optional (install_models.sh --cuda), with the user's own CUDA 13 and cuDNN 9 |
Microsoft's CUDA plugin execution provider 0.2.0, unchanged; CUDA/cuDNN not included | MIT |
License texts are in licenses/. Training-data terms of the original models
(for example SA-1B for Segment Anything and SAM 2, DIS5K for BiRefNet, Places2 for MI-GAN, ADE20K for the sky model, WIDER FACE for YuNet) are their authors' and are listed in
Rawtonic's docs/MODELS.md.