Hawkynt/photomanager-models

PhotoManager ONNX Models

Pre-converted ONNX models for PhotoManager — a cross-platform photo management and editing application.

These models are downloaded automatically by the app via Tools → Download detection models…. Each model is optional; the app degrades gracefully when a model is not installed.

Models

File	Size	Purpose	Source	Input	Output
	~258 MB	Colorization — state-of-the-art B&W→color (DDColor ECCV 2023, paper-tiny variant)	piddnad/DDColor via instant-high/DDColor-onnx	1×3×256×256 float32 [0,1] (grey RGB)	1×2×256×256 float32 (Lab a/b chroma)
	~934 MB	Colorization — DDColor artistic variant at 512×512, bolder color choices	Same as above	1×3×512×512 float32 [0,1]	1×2×512×512 float32
	~288 MB	JPEG artifact removal — blind removal of compression ringing, blocking, halos (FBCNN, Jiang et al. 2021)	jiaxi-jiang/FBCNN	1×3×H×W float32 [0,1] (dynamic)	1×3×H×W float32 (restored)
(+)	~60 MB	Scratch detection — neural-net detector for scratches/tears in old photos (Bringing Old Photos Back to Life)	microsoft/Bringing-Old-Photos-Back-to-Life	1×3×256×256 float32 [0,1]	1×1×256×256 float32 (probability)
	~15 MB	Sky segmentation — per-pixel sky probability mask for selective sky edits (SegFormer-B0 ADE20K)	nvidia/segformer-b0-finetuned-ade-512-512	1×3×H×W float32 (ImageNet-normalized)	1×150×H/4×W/4 float32 (150 ADE20K classes, sky = class 2)

Other models used by PhotoManager

These are hosted on their original repos and downloaded directly:

Model	Purpose	Source
YOLOv8n	Object detection / auto-keyword	Hyuto/yolov8-onnxruntime-web
UltraFace RFB-320	Face detection	onnx/models
ArcFace	Face embedding (512-D)	onnx/models
MODNet	Subject/portrait segmentation	Xenova/modnet
NAFNet variants	AI denoising (SIDD/GoPro)	deepghs/image_restoration
SCUNet-GAN	AI denoising (scan/old photo)	deepghs/image_restoration
RealESRGAN-x4	AI upscaling (4×)	imgdesignart/realesrgan-x4-onnx
SwinIR-x4	AI upscaling (4×, alternative)	rocca/swin-ir-onnx
DeOldify	Colorization (legacy)	thookham/DeOldify
GFPGAN v1.4	Face restoration	Meeperomi/GFPGANv1.4-onnx
LaMa-Fourier	Inpainting	opencv/inpainting_lama
SegFormer-B0 ADE150	Multi-class semantic segmentation (grass, water, road, building, person, sand, rock, foliage, snow, mountain — 150 classes)	nvidia/segformer-b0-finetuned-ade-512-512
Depth Anything V2 small	Monocular depth estimation (portrait bokeh, depth-aware masks)	onnx-community/depth-anything-v2-small
Zero-DCE++	Low-light enhancement (~52 KB)	Li-Chongyi/Zero-DCE_extension
AOD-Net	Image dehazing (~9 KB)	weberwcwei/AODnet-by-pytorch
NIMA MobileNetV2	Aesthetic quality scoring (1.0-10.0)	truskovskiyk/nima.pytorch
SigLIP (vision+text)	Auto-keyword tagging via CLIP	Hosted here

Conversion notes

• FBCNN: Exported from the official PyTorch checkpoint (jiaxi-jiang/FBCNN, nb=4) using torch.onnx.export with opset 14 and dynamic H/W axes.
• Sky segmenter & SegFormer-B0 ADE150 (sky-segmenter.onnx / segformer-ade150.onnx): Both exported from nvidia/segformer-b0-finetuned-ade-512-512 via torch.onnx.export (opset 17, dynamic N/H/W axes). Output shape [N, 150, H/4, W/4] logits. OnnxSkySegmenter and OnnxAdeSegmenter apply per-pixel argmax across the 150 channels; the consumer picks any class index (sky=2, grass=9, water=21, etc.) for binary masks.
• DDColor: Converted from the upstream DDColor-onnx project's Google Drive zip.
• BOPB scratch detector: Converted from the Microsoft BOPB repo's PyTorch weights.

License

Models are provided under their respective original licenses. This repository aggregates ONNX conversions for convenience. See each source repo for license details.