ingmebs/bspicture-modelos

Mirror of https://github.com/DuKeRd-maker/bspicture-modelos (release modelos-v2). BS Picture downloads from GitHub first and from here if GitHub is not available; every file is checked against its SHA-256.

BS Picture – AI models

Optional AI models used by BS Picture, a fast photo viewer and editor for Windows (Microsoft Store).

The models are not included in the app. BS Picture downloads a model only the first time the user chooses the feature that needs it, checks its SHA-256 fingerprint, and then runs it locally on the PC with ONNX Runtime. Photos are never uploaded anywhere.

Files

The files are published as assets of the Releases of this repository.

Release	File	Feature in BS Picture	Model	License	SHA-256
modelos-v2 (current)	quitar_fondo_v2.onnx (97 MB)	Remove background	BiRefNet lite (Swin-T), FP16 weights	MIT	E811456A9088BFF0AFC5623845C54B1C9F020EFBA8F7CA8743539D0C30E954AC
modelos-v2 (current)	borrar_objetos_v1.onnx (101 MB)	Erase objects	LaMa (big-lama), FP16 weights	Apache License 2.0	6C8A0CC49421023B3D2A0712A19C33F44BD300A001373A6E88FF3C6E1656E0F3
modelos-v2 (current)	colorear_v1.onnx (110 MB)	Colorize black and white	DDColor tiny, FP16 weights	Apache License 2.0	36BA53ED2C91D9D192A5EF5A30B4623440930B0CE2C995B8FA493A55E4DACFF0
modelos-v2 (current)	reparar_rayas_v1.onnx (72 MB)	Repair scratches (finds them; the filling uses borrar_objetos_v1.onnx)	Scratch detection network of "Bringing Old Photos Back to Life", FP16 weights	MIT	3EA1A382FB8660490F8EC44875ADAA51F563342F0B2AC203D98574AD5631528E
modelos-v2 (current)	fondos_v2.zip (53 MB)	New backgrounds (Backgrounds tab)	104 images generated with Stable Diffusion XL 1.0	see below	86CBE49E880260AEF207C7747CA7CD6086F8370BD9C22461FAD66ADFB1978F6C
modelos-v2 (older app builds)	fondos_v1.zip (50 MB)	New backgrounds (Backgrounds tab)	96 images generated with Stable Diffusion XL 1.0	see below	FF903D6609D7B66291F729F5CE850B60B04F4ABD9B111DAADD81DA4EB17E47AE
modelos-v2 (current)	ampliar_rapido_v1.onnx (5 MB)	Upscale (fast)	Real-ESRGAN realesr-general-x4v3	BSD 3-Clause	939255A67D1E756C01A4CFD23601459E3995AB43E1A462E544F73BCB84D9F28F
modelos-v2 (current)	ampliar_calidad_v1.onnx (32 MB)	Upscale (best quality)	Real-ESRGAN RealESRGAN_x4plus, FP16 weights	BSD 3-Clause	BABB8B11E85FEB1BD181C6AB8A1E0426C4FD240FF0B2ABA5B4E9972394E71EC6
modelos-v1 (older app versions)	quitar_fondo_v1.onnx (86 MB)	Remove background	IS-Net general-use (DIS), FP16	Apache License 2.0	3FD7A4F645A1D330CB275611196BBB3D6687935BC429CF7BC21BFB095236333E

Credits and licenses

BiRefNet – Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, Nicu Sebe. Bilateral Reference for High-Resolution Dichotomous Image Segmentation, CAAI Artificial Intelligence Research, 2024. https://github.com/ZhengPeng7/BiRefNet – MIT License (see LICENSE-BiRefNet-MIT.txt in the release).

IS-Net (DIS) – Xuebin Qin, Hang Dai, Xiaobin Hu, Deng-Ping Fan, Ling Shao, Luc Van Gool. Highly Accurate Dichotomous Image Segmentation, ECCV 2022. https://github.com/xuebinqin/DIS – Apache License 2.0 (see LICENSE-IS-Net-Apache-2.0.txt in the release).

The ONNX exports of the background models come from rembg by Daniel Gatis (https://github.com/danielgatis/rembg, MIT License).

LaMa – Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, Victor Lempitsky. Resolution-robust Large Mask Inpainting with Fourier Convolutions, WACV 2022. https://github.com/advimman/lama – Apache License 2.0 (see LICENSE-LaMa-Apache-2.0.txt in the release). The ONNX export comes from Carve/LaMa-ONNX (https://huggingface.co/Carve/LaMa-ONNX, Apache License 2.0).

Real-ESRGAN – Xintao Wang, Liangbin Xie, Chao Dong, Ying Shan. Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data, ICCVW 2021. https://github.com/xinntao/Real-ESRGAN – BSD 3-Clause License (see LICENSE-Real-ESRGAN-BSD-3.txt in the release).

Bringing Old Photos Back to Life (scratch detection) – Ziyu Wan, Bo Zhang, Dongdong Chen, Pan Zhang, Dong Chen, Jing Liao, Fang Wen. Bringing Old Photos Back to Life, CVPR 2020, and Old Photo Restoration via Deep Latent Space Translation, TPAMI 2022. https://github.com/microsoft/Bringing-Old-Photos-Back-to-Life – Copyright (c) Microsoft Corporation, MIT License (see LICENSE-BOPBTL-scratch-detection-MIT.txt).

DDColor – Xiaoyang Kang, Tao Yang, Wenqi Ouyang, Peiran Ren, Lingzhi Li, Xuansong Xie. DDColor: Towards Photo-Realistic Image Colorization via Dual Decoders, ICCV 2023. https://github.com/piddnad/DDColor – Apache License 2.0 (see LICENSE-DDColor-Apache-2.0.txt in the release). The ONNX export of the ddcolor_paper_tiny checkpoint comes from Qualcomm AI Hub Models (https://huggingface.co/qualcomm/DDColor, Apache License 2.0).

Themed backgrounds (fondos_v2.zip, 104 images; fondos_v1.zip, the first 96) – original images (garden, beach, party, city, nature, sky, indoors, holidays, studio, abstract, wedding, graduation, sports, office, kids, travel) generated for BS Picture with Stable Diffusion XL 1.0 by Stability AI (https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0, CreativeML Open RAIL++-M License, which places no restrictions on the use of generated images other than those of its use-based restrictions) and upscaled ×2 with Real-ESRGAN x4plus. They contain no people, logos or text. You may use them freely together with BS Picture.

Modifications:

• reparar_rayas_v1.onnx: built from the official checkpoints/detection/FT_Epoch_latest.pt (only the network weights; the training state was left out). Each batch normalization was folded into the preceding convolution and the weights are stored in FP16 and cast back to 32-bit when the model is loaded; computation stays in FP32. Its output matches the original PyTorch network. Input: grayscale photo scaled to -1..1 (shorter side 256 px, sides multiple of 16); output: logits (sigmoid ≥ 0.4 is a scratch).

• colorear_v1.onnx: the external weight file of the ONNX export was merged into a single file, with the weights stored in FP16 and cast back to 32-bit when the model is loaded; computation stays in FP32.

• ampliar_rapido_v1.onnx and ampliar_calidad_v1.onnx: converted to ONNX from the official realesr-general-x4v3.pth and RealESRGAN_x4plus.pth weights of the Real-ESRGAN releases (same networks, SRVGGNetCompact and RRDBNet; input RGB 0–1, output ×4). In ampliar_calidad_v1.onnx the weights are stored in FP16 and cast back to 32-bit when the model is loaded; computation stays in FP32.

• borrar_objetos_v1.onnx: the weights of the FP32 ONNX model are stored in half precision (FP16) and cast back to 32-bit when the model is loaded, so all computation stays in FP32 (a full FP16 conversion breaks the model's Fourier convolutions). The network was not otherwise changed.

• quitar_fondo_v2.onnx: in the ONNX export, every deformable convolution of the decoder had been expanded into gather/multiply operations that need about 6 GB of memory at 1024 × 1024. They were rewritten with standard operators that GPUs also support (DirectML): for each kernel position, GridSample (bilinear, zero padding, aligned corners) samples the input at the offset points, the result is multiplied by the modulation mask and by that position's slice of the weights as a 1 × 1 convolution, and the partial results plus the bias are summed. This gives the same output (average difference 0.002 of 255 on the mask) with much less memory. The weights are stored in half precision (FP16) and cast back to 32-bit when the model is loaded; computation stays in FP32. The weights were not otherwise changed.

• quitar_fondo_v1.onnx: the original FP32 ONNX model was converted to half precision (FP16) with ONNX Runtime's float16 converter, keeping 32-bit inputs and outputs. The network and its weights were not otherwise changed.