CoreML-Models
April 24, 2026 · View on GitHub
Converted Core ML Model Zoo.
Core ML is a machine learning framework by Apple. If you are iOS developer, you can easly use machine learning models in your Xcode project.
Try the iOS sample-app collection (sample_apps/CoreMLModelsApp) on the App Store:
How to use
Take a look this model zoo, and if you found the CoreML model you want, download the model from google drive link and bundle it in your project. Or if the model have sample project link, try it and see how to use the model in the project. You are free to do or not.
If you like this repository, please give me a star so I can do my best.
Section Link
-
Stable Diffusion :text2image
How to get the model
You can get the model converted to CoreML format from the link of Google drive. See the section below for how to use it in Xcode. The license for each model conforms to the license for the original project.
Image Classifier
Efficientnet
| Google Drive Link | Size | Dataset | Original Project | License |
|---|---|---|---|---|
| Efficientnetb0 | 22.7 MB | ImageNet | TensorFlowHub | Apache2.0 |
Efficientnetv2
| Google Drive Link | Size | Dataset | Original Project | License | Year |
|---|---|---|---|---|---|
| Efficientnetv2 | 85.8 MB | ImageNet | Google/autoML | Apache2.0 | 2021 |
VisionTransformer
An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale.
| Google Drive Link | Size | Dataset | Original Project | License | Year |
|---|---|---|---|---|---|
| VisionTransformer-B16 | 347.5 MB | ImageNet | google-research/vision_transformer | Apache2.0 | 2021 |
Conformer
Local Features Coupling Global Representations for Visual Recognition.
| Google Drive Link | Size | Dataset | Original Project | License | Year |
|---|---|---|---|---|---|
| Conformer-tiny-p16 | 94.1 MB | ImageNet | pengzhiliang/Conformer | Apache2.0 | 2021 |
DeiT
Data-efficient Image Transformers
| Google Drive Link | Size | Dataset | Original Project | License | Year |
|---|---|---|---|---|---|
| DeiT-base384 | 350.5 MB | ImageNet | facebookresearch/deit | Apache2.0 | 2021 |
RepVGG
Making VGG-style ConvNets Great Again
| Google Drive Link | Size | Dataset | Original Project | License | Year |
|---|---|---|---|---|---|
| RepVGG-A0 | 33.3 MB | ImageNet | DingXiaoH/RepVGG | MIT | 2021 |
RegNet
Designing Network Design Spaces
| Google Drive Link | Size | Dataset | Original Project | License | Year |
|---|---|---|---|---|---|
| regnet_y_400mf | 16.5 MB | ImageNet | TORCHVISION.MODELS | MIT | 2020 |
MobileViTv2
CVNets: A library for training computer vision networks
| Google Drive Link | Size | Dataset | Original Project | License | Year | Conversion Script |
|---|---|---|---|---|---|---|
| MobileViTv2 | 18.8 MB | ImageNet | apple/ml-cvnets | apple | 2022 |  |
Object Detection
D-FINE
| Download Link | Size | Output | Original Project | License | Note | Sample Project |
|---|---|---|---|---|---|---|
| dfine-n-coco | 13MB | Confidence(MultiArray (Float32 300 × 80)), Coordinates (MultiArray (Float32 300 × 4)) | Peterande/D-FINE | Apache 2.0 | Input 640×640. Coordinates are normalized cxcywh. No NMS — filter by confidence threshold. | peaceofcake DFINEDemo |
RF-DETR
| Download Link | Size | Output | Original Project | License | Note | Sample Project |
|---|---|---|---|---|---|---|
| rfdetr-n-coco | 95MB | Confidence(MultiArray (Float32 300 × 91)), Coordinates (MultiArray (Float32 300 × 4)) | roboflow/rf-detr | Apache 2.0 | Input 384×384. 91 classes (index 0 = background, 1-90 = COCO category IDs). Coordinates are normalized cxcywh. No NMS. | peaceofcake DFINEDemo |
YOLOv5s
| Google Drive Link | Size | Output | Original Project | License | Note | Sample Project |
|---|---|---|---|---|---|---|
| YOLOv5s | 29.3MB | Confidence(MultiArray (Double 0 × 80)), Coordinates (MultiArray (Double 0 × 4)) | ultralytics/yolov5 | GNU | Non Maximum Suppression has been added. | CoreML-YOLOv5 |
YOLOv7
| Google Drive Link | Size | Output | Original Project | License | Note | Sample Project | Conversion Script |
|---|---|---|---|---|---|---|---|
| YOLOv7 | 147.9MB | Confidence(MultiArray (Double 0 × 80)), Coordinates (MultiArray (Double 0 × 4)) | WongKinYiu/yolov7 | GNU | Non Maximum Suppression has been added. | CoreML-YOLOv5 | |
YOLOv8
| Google Drive Link | Size | Output | Original Project | License | Note | Sample Project |
|---|---|---|---|---|---|---|
| YOLOv8s | 45.1MB | Confidence(MultiArray (Double 0 × 80)), Coordinates (MultiArray (Double 0 × 4)) | ultralytics/ultralytics | GNU | Non Maximum Suppression has been added. | CoreML-YOLOv5 |
YOLOv9
YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information. Uses PGI and GELAN architecture for efficient object detection.
| Download Link | Size | Output | Original Project | License | Year | Note | Sample Project |
|---|---|---|---|---|---|---|---|
| yolov9s.mlpackage.zip | 14 MB | Confidence (MultiArray (Double 0 × 80)), Coordinates (MultiArray (Double 0 × 4)) | WongKinYiu/yolov9 | GPL-3.0 | 2024 | Non Maximum Suppression has been added. | YOLOv9Demo |
YOLOv10
YOLOv10: Real-Time End-to-End Object Detection. NMS-free architecture using consistent dual assignments — no post-processing needed.
| Download Link | Size | Output | Original Project | License | Year | Note | Sample Project |
|---|---|---|---|---|---|---|---|
| yolov10s.mlpackage.zip | 14 MB | MultiArray (1 × 300 × 6) | THU-MIG/yolov10 | AGPL-3.0 | 2024 | NMS-free end-to-end detection. | YOLO26Demo |
YOLO11
YOLO11: Ultralytics latest YOLO with improved backbone and neck architecture. 22% fewer parameters than YOLOv8 with higher mAP.
| Download Link | Size | Output | Original Project | License | Year | Note | Sample Project |
|---|---|---|---|---|---|---|---|
| yolo11s.mlpackage.zip | 18 MB | Confidence (MultiArray (Double 0 × 80)), Coordinates (MultiArray (Double 0 × 4)) | ultralytics/ultralytics | AGPL-3.0 | 2024 | Non Maximum Suppression has been added. | YOLOv9Demo |
YOLO26
YOLO26: Edge-first vision AI with NMS-free end-to-end detection. Up to 43% faster CPU inference vs YOLO11 with DFL removal and ProgLoss.
| Download Link | Size | Output | Original Project | License | Year | Note | Sample Project |
|---|---|---|---|---|---|---|---|
| yolo26s.mlpackage.zip | 18 MB | MultiArray (1 × 300 × 6) | ultralytics/ultralytics | AGPL-3.0 | 2026 | NMS-free end-to-end detection. | YOLO26Demo |
YOLO-World
YOLO-World: Real-Time Open-Vocabulary Object Detection. Type any text query and detect it — no fixed class list. Uses CLIP text encoder for open-vocabulary matching.
| Download Link | Size | Description | Original Project | License | Year | Sample Project |
|---|---|---|---|---|---|---|
| yoloworld_detector.mlpackage.zip | 25 MB | YOLO-World V2-S visual detector | AILab-CVC/YOLO-World | GPL-3.0 | 2024 | YOLOWorldDemo |
| clip_text_encoder.mlpackage.zip | 121 MB | CLIP ViT-B/32 text encoder | openai/CLIP | MIT | 2021 | — |
| clip_vocab.json.zip | 1.6 MB | BPE vocabulary for tokenizer | — | — | — | — |
Multi-Object Tracking
ByteTrack
ByteTrack: Multi-Object Tracking by Associating Every Detection Box. Pure-Swift on-device tracker that adds persistent IDs to any of the object detectors above. Pairs a per-track 8D constant-velocity Kalman filter with a two-stage IoU association — high-confidence detections are matched first, then low-confidence detections are used to rescue tracks about to be lost through motion blur and brief occlusions. No appearance / ReID network, so it runs for free on top of an existing detector.
| Implementation | Source | Paper | License | Year | Note | Sample Project |
|---|---|---|---|---|---|---|
| Pure Swift (no download) | Tracker.swift | ByteTrack (arXiv 2110.06864) | MIT (this port) / Original | 2022 | 8D Kalman + two-stage IoU association, class-aware, greedy matching, lost-track buffer of 30 frames. Drop-in on top of any [Detection] stream. | YOLO26Demo |
Segmentation
U2Net
| Google Drive Link | Size | Output | Original Project | License |
|---|---|---|---|---|
| U2Net | 175.9 MB | Image(GRAYSCALE 320 × 320) | xuebinqin/U-2-Net | Apache |
| U2Netp | 4.6 MB | Image(GRAYSCALE 320 × 320) | xuebinqin/U-2-Net | Apache |
IS-Net
| Google Drive Link | Size | Output | Original Project | License | Year | Conversion Script |
|---|---|---|---|---|---|---|
| IS-Net | 176.1 MB | Image(GRAYSCALE 1024 × 1024) | xuebinqin/DIS | Apache | 2022 | |
| IS-Net-General-Use | 176.1 MB | Image(GRAYSCALE 1024 × 1024) | xuebinqin/DIS | Apache | 2022 | |
RMBG1.4
RMBG1.4 - The IS-Net enhanced with our unique training scheme and proprietary dataset.
| Download Link | Size | Output | Original Project | License | year | Sample Project | Conversion Script |
|---|---|---|---|---|---|---|---|
| RMBG_1_4.mlpackage.zip | 42 MB (INT8) | Alpha mask 1024x1024 | briaai/RMBG-1.4 | Creative Commons | 2024 | RMBGDemo | convert_rmbg.py |
face-Parsing
| Google Drive Link | Size | Output | Original Project | License | Sample Project |
|---|---|---|---|---|---|
| face-Parsing | 53.2 MB | MultiArray(1 x 512 × 512) | zllrunning/face-parsing.PyTorch | MIT | CoreML-face-parsing |
Segformer
Simple and Efficient Design for Semantic Segmentation with Transformers
| Google Drive Link | Size | Output | Original Project | License | year |
|---|---|---|---|---|---|
| SegFormer_mit-b0_1024x1024_cityscapes | 14.9 MB | MultiArray(512 × 1024) | NVlabs/SegFormer | NVIDIA | 2021 |
BiSeNetV2
Bilateral Network with Guided Aggregation for Real-time Semantic Segmentation
| Google Drive Link | Size | Output | Original Project | License | year |
|---|---|---|---|---|---|
| BiSeNetV2_1024x1024_cityscapes | 12.8 MB | MultiArray | ycszen/BiSeNet | Apache2.0 | 2021 |
DNL
Disentangled Non-Local Neural Networks
| Google Drive Link | Size | Output | Dataset | Original Project | License | year |
|---|---|---|---|---|---|---|
| dnl_r50-d8_512x512_80k_ade20k | 190.8 MB | MultiArray[512x512] | ADE20K | yinmh17/DNL-Semantic-Segmentation | Apache2.0 | 2020 |
ISANet
Interlaced Sparse Self-Attention for Semantic Segmentation
| Google Drive Link | Size | Output | Dataset | Original Project | License | year |
|---|---|---|---|---|---|---|
| isanet_r50-d8_512x512_80k_ade20k | 141.5 MB | MultiArray[512x512] | ADE20K | openseg-group/openseg.pytorch | MIT | ArXiv'2019/IJCV'2021 |
FastFCN
Rethinking Dilated Convolution in the Backbone for Semantic Segmentation
| Google Drive Link | Size | Output | Dataset | Original Project | License | year |
|---|---|---|---|---|---|---|
| fastfcn_r50-d32_jpu_aspp_512x512_80k_ade20k | 326.2 MB | MultiArray[512x512] | ADE20K | wuhuikai/FastFCN | MIT | ArXiv'2019 |
GCNet
Non-local Networks Meet Squeeze-Excitation Networks and Beyond
| Google Drive Link | Size | Output | Dataset | Original Project | License | year |
|---|---|---|---|---|---|---|
| gcnet_r50-d8_512x512_20k_voc12aug | 189 MB | MultiArray[512x512] | PascalVOC | xvjiarui/GCNet | Apache License 2.0 | ICCVW'2019/TPAMI'2020 |
DANet
Dual Attention Network for Scene Segmentation(CVPR2019)
| Google Drive Link | Size | Output | Dataset | Original Project | License | year |
|---|---|---|---|---|---|---|
| danet_r50-d8_512x1024_40k_cityscapes | 189.7 MB | MultiArray[512x1024] | CityScapes | junfu1115/DANet | MIT | CVPR2019 |
Semantic-FPN
Panoptic Feature Pyramid Networks
| Google Drive Link | Size | Output | Dataset | Original Project | License | year |
|---|---|---|---|---|---|---|
| fpn_r50_512x1024_80k_cityscapes | 108.6 MB | MultiArray[512x1024] | CityScapes | facebookresearch/detectron2 | Apache License 2.0 | 2019 |
cloths_segmentation
Code for binary segmentation of various cloths.
| Google Drive Link | Size | Output | Dataset | Original Project | License | year |
|---|---|---|---|---|---|---|
| clothSegmentation | 50.1 MB | Image(GrayScale 640x960) | fashion-2019-FGVC6 | facebookresearch/detectron2 | MIT | 2020 |
easyportrait
EasyPortrait - Face Parsing and Portrait Segmentation Dataset.
| Google Drive Link | Size | Output | Original Project | License | year | Swift sample | Conversion Script |
|---|---|---|---|---|---|---|---|
| easyportrait-segformer512-fp | 7.6 MB | Image(GrayScale 512x512) * 9 | hukenovs/easyportrait | Creative Commons | 2023 | easyportrait-coreml | |
MobileSAM
Faster Segment Anything: Towards Lightweight SAM for Mobile Applications. MobileSAM replaces the heavy ViT-H image encoder with a lightweight ViT-Tiny encoder via decoupled knowledge distillation, making it ~60x smaller and ~40x faster than the original SAM.
| Download Link | Size | Output | Original Project | License | Year | Sample Project |
|---|---|---|---|---|---|---|
| MobileSAM.zip | 23 MB (Encoder 13 MB + Decoder 9.8 MB) | Segmentation Mask | ChaoningZhang/MobileSAM | Apache 2.0 | 2023 | SamKit |
SAM2-Tiny
SAM 2: Segment Anything in Images and Videos. SAM 2 extends promptable segmentation from images to videos using a streaming architecture with memory. The Tiny variant uses a Hiera-T backbone for efficient on-device inference.
| Download Link | Size | Output | Original Project | License | Year | Sample Project |
|---|---|---|---|---|---|---|
| SAM2Tiny.zip | 76 MB (ImageEncoder 64 MB + PromptEncoder 2 MB + MaskDecoder 9.8 MB) | Segmentation Mask | facebookresearch/sam2 | Apache 2.0 | 2024 | SamKit |
Video Matting
MatAnyone
pq-yang/MatAnyone (CVPR 2025) — temporally consistent video matting with object-level memory propagation. Given a first-frame mask the network tracks and refines an alpha matte across the whole clip, holding sharp edges (hair, semitransparent regions) much better than per-frame matting baselines. Built on the Cutie video object segmentation backbone with a dedicated mask decoder for matting.
The CoreML port splits the network into 5 stateless modules so the per-frame memory state machine can live in Swift while CoreML handles the heavy compute. End-to-end alpha matte parity vs the official PyTorch reference: MAE < 2e-4, correlation 0.9999+ across 18 frames including 3 memory cycles.
The sample app uses Vision's VNGeneratePersonSegmentationRequest to bootstrap the first-frame mask automatically — pick a video, tap "Remove BG", and it composites the foreground over the chosen background colour.
| Download Link | Size | Input | Output | Original Project | License | Year | Sample Project | Conversion Script |
|---|---|---|---|---|---|---|---|---|
| MatAnyone (5 mlpackages, ~111 MB FP16 total) | 111 MB | image [1,3,432,768] (per-frame state in Swift) | alpha matte [1,1,432,768] | pq-yang/MatAnyone | NTU S-Lab 1.0 | 2025 | MatAnyoneDemo | convert_matanyone.py |
See sample_apps/MatAnyoneDemo/README.md for the per-frame state machine, the 5-module split, and conversion details.
Super Resolution
Real ESRGAN
| Google Drive Link | Size | Output | Original Project | License | year |
|---|---|---|---|---|---|
| Real ESRGAN4x | 66.9 MB | Image(RGB 2048x2048) | xinntao/Real-ESRGAN | BSD 3-Clause License | 2021 |
| Real ESRGAN Anime4x | 66.9 MB | Image(RGB 2048x2048) | xinntao/Real-ESRGAN | BSD 3-Clause License | 2021 |
GFPGAN
Towards Real-World Blind Face Restoration with Generative Facial Prior
| Google Drive Link | Size | Output | Original Project | License | year |
|---|---|---|---|---|---|
| GFPGAN | 337.4 MB | Image(RGB 512x512) | TencentARC/GFPGAN | Apache2.0 | 2021 |
BSRGAN
| Google Drive Link | Size | Output | Original Project | License | year |
|---|---|---|---|---|---|
| BSRGAN | 66.9 MB | Image(RGB 2048x2048) | cszn/BSRGAN | 2021 |
A-ESRGAN
| Google Drive Link | Size | Output | Original Project | License | year | Conversion Script |
|---|---|---|---|---|---|---|
| A-ESRGAN | 63.8 MB | Image(RGB 1024x1024) | aesrgan/A-ESRGANN | BSD 3-Clause License | 2021 | |
Beby-GAN
Best-Buddy GANs for Highly Detailed Image Super-Resolution
| Google Drive Link | Size | Output | Original Project | License | year |
|---|---|---|---|---|---|
| Beby-GAN | 66.9 MB | Image(RGB 2048x2048) | dvlab-research/Simple-SR | MIT | 2021 |
RRDN
The Residual in Residual Dense Network for image super-scaling.
| Google Drive Link | Size | Output | Original Project | License | year |
|---|---|---|---|---|---|
| RRDN | 16.8 MB | Image(RGB 2048x2048) | idealo/image-super-resolution | Apache2.0 | 2018 |
Fast-SRGAN
Fast-SRGAN.
| Google Drive Link | Size | Output | Original Project | License | year |
|---|---|---|---|---|---|
| Fast-SRGAN | 628 KB | Image(RGB 1024x1024) | HasnainRaz/Fast-SRGAN | MIT | 2019 |
ESRGAN
Enhanced-SRGAN.
| Google Drive Link | Size | Output | Original Project | License | year |
|---|---|---|---|---|---|
| ESRGAN | 66.9 MB | Image(RGB 2048x2048) | xinntao/ESRGAN | Apache 2.0 | 2018 |
UltraSharp
Pretrained: 4xESRGAN
| Google Drive Link | Size | Output | Original Project | License | year |
|---|---|---|---|---|---|
| UltraSharp | 34 MB | Image(RGB 1024x1024) | Kim2019/ | CC-BY-NC-SA-4.0 | 2021 |
SRGAN
Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network.
| Google Drive Link | Size | Output | Original Project | License | year |
|---|---|---|---|---|---|
| SRGAN | 6.1 MB | Image(RGB 2048x2048) | dongheehand/SRGAN-PyTorch | 2017 |
SRResNet
Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network.
| Google Drive Link | Size | Output | Original Project | License | year |
|---|---|---|---|---|---|
| SRResNet | 6.1 MB | Image(RGB 2048x2048) | dongheehand/SRGAN-PyTorch | 2017 |
LESRCNN
Lightweight Image Super-Resolution with Enhanced CNN.
| Google Drive Link | Size | Output | Original Project | License | year | Conversion Script |
|---|---|---|---|---|---|---|
| LESRCNN | 4.3 MB | Image(RGB 512x512) | hellloxiaotian/LESRCNN | 2020 | |
MMRealSR
Metric Learning based Interactive Modulation for Real-World Super-Resolution
| Google Drive Link | Size | Output | Original Project | License | year | Conversion Script |
|---|---|---|---|---|---|---|
| MMRealSRGAN | 104.6 MB | Image(RGB 1024x1024) | TencentARC/MM-RealSR | BSD 3-Clause | 2022 | |
| MMRealSRNet | 104.6 MB | Image(RGB 1024x1024) | TencentARC/MM-RealSR | BSD 3-Clause | 2022 | |
DASR
Pytorch implementation of "Unsupervised Degradation Representation Learning for Blind Super-Resolution", CVPR 2021
| Google Drive Link | Size | Output | Original Project | License | year |
|---|---|---|---|---|---|
| DASR | 12.1 MB | Image(RGB 1024x1024) | The-Learning-And-Vision-Atelier-LAVA/DASR | MIT | 2022 |
SinSR
wyf0912/SinSR — single-step diffusion-based super-resolution (CVPR 2024, ~113M params). Distilled from ResShift for one-step 4x upscaling. Uses a Swin Transformer UNet with VQ-VAE latent space.
Left: bicubic 4x upscale, Right: SinSR single-step diffusion SR (128x128 → 512x512)
3 CoreML models: VQ-VAE encoder, Swin-UNet denoiser (single step), and VQ-VAE decoder with vector quantization.
| Download Link | Size | Input | Output | Original Project | License | Year | Sample Project | Conversion Script |
|---|---|---|---|---|---|---|---|---|
| SinSR_Encoder.mlpackage.zip | 39 MB | image [1,3,1024,1024] | latent [1,3,256,256] | wyf0912/SinSR | S-Lab | 2024 | SinSRDemo | convert_sinsr.py |
| SinSR_Denoiser.mlpackage.zip | 420 MB | input [1,6,256,256] | predicted_latent [1,3,256,256] | |||||
| SinSR_Decoder.mlpackage.zip | 58 MB | latent [1,3,256,256] | image [1,3,1024,1024] |
See sample_apps/SinSRDemo/README.md for the inference pipeline and conversion details.
Low Light Enhancement
StableLLVE
Learning Temporal Consistency for Low Light Video Enhancement from Single Images.
| Google Drive Link | Size | Output | Original Project | License | Year |
|---|---|---|---|---|---|
| StableLLVE | 17.3 MB | Image(RGB 512x512) | zkawfanx/StableLLVE | MIT | 2021 |
Zero-DCE
Zero-Reference Deep Curve Estimation for Low-Light Image Enhancement
| Google Drive Link | Size | Output | Original Project | License | Year | Conversion Script |
|---|---|---|---|---|---|---|
| Zero-DCE | 320KB | Image(RGB 512x512) | Li-Chongyi/Zero-DCE | See Repo | 2021 | |
Retinexformer
Retinexformer: One-stage Retinex-based Transformer for Low-light Image Enhancement
| Google Drive Link | Size | Output | Original Project | License | Year | Conversion Script |
|---|---|---|---|---|---|---|
| ZRetinexformer FiveK | 3.4MB | Image(RGB 512x512) | caiyuanhao1998/Retinexformer | MIT | 2023 | |
| ZRetinexformer NTIRE | 3.4MB | Image(RGB 512x512) | caiyuanhao1998/Retinexformer | MIT | 2023 | |
Image Restoration
MPRNet
Multi-Stage Progressive Image Restoration.
Debluring
Denoising
Deraining
| Google Drive Link | Size | Output | Original Project | License | Year |
|---|---|---|---|---|---|
| MPRNetDebluring | 137.1 MB | Image(RGB 512x512) | swz30/MPRNet | MIT | 2021 |
| MPRNetDeNoising | 108 MB | Image(RGB 512x512) | swz30/MPRNet | MIT | 2021 |
| MPRNetDeraining | 24.5 MB | Image(RGB 512x512) | swz30/MPRNet | MIT | 2021 |
MIRNetv2
Learning Enriched Features for Fast Image Restoration and Enhancement.
Denoising
Super Resolution
Contrast Enhancement
Low Light Enhancement
| Google Drive Link | Size | Output | Original Project | License | Year | Conversion Script |
|---|---|---|---|---|---|---|
| MIRNetv2Denoising | 42.5 MB | Image(RGB 512x512) | swz30/MIRNetv2 | ACADEMIC PUBLIC LICENSE | 2022 | |
| MIRNetv2SuperResolution | 42.5 MB | Image(RGB 512x512) | swz30/MIRNetv2 | ACADEMIC PUBLIC LICENSE | 2022 | |
| MIRNetv2ContrastEnhancement | 42.5 MB | Image(RGB 512x512) | swz30/MIRNetv2 | ACADEMIC PUBLIC LICENSE | 2022 | |
| MIRNetv2LowLightEnhancement | 42.5 MB | Image(RGB 512x512) | swz30/MIRNetv2 | ACADEMIC PUBLIC LICENSE | 2022 | |
Image Generation
MobileStyleGAN
| Google Drive Link | Size | Output | Original Project | License | Sample Project |
|---|---|---|---|---|---|
| MobileStyleGAN | 38.6MB | Image(Color 1024 × 1024) | bes-dev/MobileStyleGAN.pytorch | Nvidia Source Code License-NC | CoreML-StyleGAN |
DCGAN
| Google Drive Link | Size | Output | Original Project |
|---|---|---|---|
| DCGAN | 9.2MB | MultiArray | TensorFlowCore |
Image2Image
Anime2Sketch
| Google Drive Link | Size | Output | Original Project | License | Usage |
|---|---|---|---|---|---|
| Anime2Sketch | 217.7MB | Image(Color 512 × 512) | Mukosame/Anime2Sketch | MIT | Drop an image to preview |
AnimeGAN2Face_Paint_512_v2
| Google Drive Link | Size | Output | Original Project | Conversion Script |
|---|---|---|---|---|
| AnimeGAN2Face_Paint_512_v2 | 8.6MB | Image(Color 512 × 512) | bryandlee/animegan2-pytorch | |
Photo2Cartoon
| Google Drive Link | Size | Output | Original Project | License | Note |
|---|---|---|---|---|---|
| Photo2Cartoon | 15.2 MB | Image(Color 256 × 256) | minivision-ai/photo2cartoon | MIT | The output is little bit different from the original model. It cause some operations were converted replaced manually. |
AnimeGANv2_Hayao
| Google Drive Link | Size | Output | Original Project | Sample |
|---|---|---|---|---|
| AnimeGANv2_Hayao | 8.7MB | Image(256 x 256) | TachibanaYoshino/AnimeGANv2 | AnimeGANv2-iOS |
AnimeGANv2_Paprika
| Google Drive Link | Size | Output | Original Project |
|---|---|---|---|
| AnimeGANv2_Paprika | 8.7MB | Image(256 x 256) | TachibanaYoshino/AnimeGANv2 |
WarpGAN Caricature
| Google Drive Link | Size | Output | Original Project |
|---|---|---|---|
| WarpGAN Caricature | 35.5MB | Image(256 x 256) | seasonSH/WarpGAN |
UGATIT_selfie2anime
| Google Drive Link | Size | Output | Original Project |
|---|---|---|---|
| UGATIT_selfie2anime | 266.2MB(quantized) | Image(256x256) | taki0112/UGATIT |
CartoonGAN
| Google Drive Link | Size | Output | Original Project |
|---|---|---|---|
| CartoonGAN_Shinkai | 44.6MB | MultiArray | mnicnc404/CartoonGan-tensorflow |
| CartoonGAN_Hayao | 44.6MB | MultiArray | mnicnc404/CartoonGan-tensorflow |
| CartoonGAN_Hosoda | 44.6MB | MultiArray | mnicnc404/CartoonGan-tensorflow |
| CartoonGAN_Paprika | 44.6MB | MultiArray | mnicnc404/CartoonGan-tensorflow |
Fast-Neural-Style-Transfer
| Google Drive Link | Size | Output | Original Project | License | Year |
|---|---|---|---|---|---|
| fast-neural-style-transfer-cuphead | 6.4MB | Image(RGB 960x640) | eriklindernoren/Fast-Neural-Style-Transfer | MIT | 2019 |
| fast-neural-style-transfer-starry-night | 6.4MB | Image(RGB 960x640) | eriklindernoren/Fast-Neural-Style-Transfer | MIT | 2019 |
| fast-neural-style-transfer-mosaic | 6.4MB | Image(RGB 960x640) | eriklindernoren/Fast-Neural-Style-Transfer | MIT | 2019 |
White_box_Cartoonization
Learning to Cartoonize Using White-box Cartoon Representations
| Google Drive Link | Size | Output | Original Project | License | Year |
|---|---|---|---|---|---|
| White_box_Cartoonization | 5.9MB | Image(1536x1536) | SystemErrorWang/White-box-Cartoonization | creativecommons | CVPR2020 |
FacialCartoonization
White-box facial image cartoonizaiton
| Google Drive Link | Size | Output | Original Project | License | Year |
|---|---|---|---|---|---|
| FacialCartoonization | 8.4MB | Image(256x256) | SystemErrorWang/FacialCartoonization | creativecommons | 2020 |
Inpainting
AOT-GAN-for-Inpainting
| Google Drive Link | Size | Output | Original Project | License | Note | Sample Project |
|---|---|---|---|---|---|---|
| AOT-GAN-for-Inpainting | 60.8MB | MLMultiArray(3,512,512) | researchmm/AOT-GAN-for-Inpainting | Apache2.0 | To use see sample. | john-rocky/Inpainting-CoreML |
Lama
| Google Drive Link | Size | Input | Output | Original Project | License | Note | Sample Project | Conversion Script |
|---|---|---|---|---|---|---|---|---|
| Lama | 216.6MB | Image (Color 800 × 800), Image (GrayScale 800 × 800) | Image (Color 800 × 800) | advimman/lama | Apache2.0 | To use see sample. | john-rocky/lama-cleaner-iOS | mallman/CoreMLaMa |
Monocular Depth Estimation
Depth Anything 3
ByteDance-Seed/Depth-Anything-3 (ICLR 2026 Oral) — relative monocular depth from a single image. DA3 Main Series uses a plain DINOv2 ViT backbone plus a DualDPT head with a unified depth-ray representation; this Core ML port exposes only the monocular depth + confidence subgraph (camera / multi-view / sky / 3DGS branches are stripped). First public Core ML conversion of DA3.
| Module | Size | Input | Output | Original Project | License | Year | Sample Project | Conversion Script |
|---|---|---|---|---|---|---|---|---|
| DA3 Small 504×504 | ~44 MB FP16 | Image (RGB 504 × 504) | depth + confidence | ByteDance-Seed/Depth-Anything-3 | Apache 2.0 | 2025 | Hub App | convert_depth_anything_v3.py |
| DA3 Base 504×504 | ~173 MB FP16 | Image (RGB 504 × 504) | depth + confidence | ByteDance-Seed/Depth-Anything-3 | Apache 2.0 | 2025 | Hub App | convert_depth_anything_v3.py |
MoGe-2
microsoft/MoGe (CVPR 2025 Oral) — open-domain monocular 3D geometry from a single image. Predicts a metric depth map, surface normals, and a confidence mask in one forward pass on a DINOv2 ViT-B backbone with three task heads. The successor to MiDaS-style relative depth: depth comes out in real meters.
Left: original photo, center: metric depth (turbo colormap), right: surface normals.
| Module | Size | Input | Output | Original Project | License | Year | Sample Project | Conversion Script |
|---|---|---|---|---|---|---|---|---|
| MoGe-2 ViT-B + normal | ~200 MB FP16 | Image (RGB 504 × 504) | depth + normal + mask + metric_scale | microsoft/MoGe | MIT | 2025 | MoGe2Demo | convert_moge2.py |
MiDaS
Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-shot Cross-dataset Transfer
| Google Drive Link | Size | Output | Original Project | License | Year | Conversion Script |
|---|---|---|---|---|---|---|
| MiDaS_Small | 66.3MB | MultiArray(1x256x256) | isl-org/MiDaS | MIT | 2022 | |
Stable Diffusion
Nitro-E
amd/Nitro-E — AMD's 304M-parameter E-MMDiT text-to-image model released October 2025. 4-step distilled variant generates 512×512 images from a prompt in ~2–3 seconds on iPhone 15+. Uses Llama 3.2 1B as the text encoder, a DC-AE f32c32 VAE decoder, and an ASA-based (Alternating Subregion Attention) diffusion transformer. Full pipeline fits in ~1.04 GB after INT4 / INT8 palettization (TextEncoder 590 MB + E-MMDiT 295 MB + VAE 159 MB).
4-step generation on iPhone, 512×512. Prompt: "a hot air balloon in the shape of a heart, grand canyon".
3 CoreML models total:
| Download Link | Size | Input | Output | Original Project | License | Year | Sample Project | Conversion Script |
|---|---|---|---|---|---|---|---|---|
| NitroE_TextEncoder.mlpackage | 590 MB (INT4) / 2.3 GB (FP16) | input_ids [1,128], attention_mask [1,128] | last_hidden_state [1,128,2048] | meta-llama/Llama-3.2-1B | Llama 3.2 (gated) | 2024 | NitroEDemo | convert_nitro_e_text_encoder.py |
| NitroE_EMMDiT.mlpackage | 295 MB (INT8) / 578 MB (FP16) | latent [1,32,16,16], encoder_hs [1,128,2048], timestep [1] | noise_pred [1,32,16,16] | amd/Nitro-E | MIT | 2025 | convert_nitro_e_emmdit.py | |
| NitroE_VAEDecoder.mlpackage | 159 MB (INT8) / 608 MB (FP32) | latent [1,32,16,16] | image [1,3,512,512] | mit-han-lab/dc-ae-f32c32-sana-1.0-diffusers | MIT | 2024 | convert_nitro_e_vae_decoder.py |
See sample_apps/NitroEDemo/README.md for the Swift FlowMatchEulerScheduler port, tokenizer details, and iOS 18 palettization notes.
Hyper-SD
ByteDance/Hyper-SD — single-step text-to-image distilled from SD1.5 via Trajectory Segmented Consistency Distillation. ByteDance reports user preference 2x over SD-Turbo at 1 step. Combined with Apple's ml-stable-diffusion (Split-Einsum attention, chunked UNet, 6-bit palettization), runs at acceptable speed and quality on iPhone 15+.
1-step generations on iPhone, 512×512. Prompts: cat with sunglasses, cyberpunk city, japanese garden, astronaut on horse.
4 CoreML models (~947 MB total): CLIP text encoder + Swin-style chunked UNet (6-bit palettized) + VAE decoder. Uses TCD scheduler for single-step inference.
| Download Link | Size | Input | Output | Original Project | License | Year | Sample Project | Conversion Script |
|---|---|---|---|---|---|---|---|---|
| HyperSDTextEncoder.mlpackage.zip | 235 MB | input_ids [1,77] | encoder_hidden_states [1,77,768] | ByteDance/Hyper-SD | OpenRAIL++ | 2024 | HyperSDDemo | convert_hypersd.py |
| HyperSDUnetChunk1.mlpackage.zip | 318 MB | latent + encoder_hs + timestep | first half intermediates | |||||
| HyperSDUnetChunk2.mlpackage.zip | 299 MB | first half outputs + skip connections | noise_pred [2,4,64,64] | |||||
| HyperSDVAEDecoder.mlpackage.zip | 95 MB | latent [1,4,64,64] | image [1,3,512,512] |
See sample_apps/HyperSDDemo/README.md for the LoRA fusion, chunked-UNet palettization, and TCD scheduler details.
stable-diffusion-v1-5
| Google Drive Link | Original Model | Original Project | License | Run on mac | Conversion Script | Year |
|---|---|---|---|---|---|---|
| stable-diffusion-v1-5 | runwayml/stable-diffusion-v1-5 | runwayml/stable-diffusion | Open RAIL M license | godly-devotion/MochiDiffusion | godly-devotion/MochiDiffusion | 2022 |
pastel-mix
Pastel Mix - a stylized latent diffusion model.This model is intended to produce high-quality, highly detailed anime style with just a few prompts.
| Google Drive Link | Original Model | License | Run on mac | Conversion Script | Year |
|---|---|---|---|---|---|
| pastelMixStylizedAnime_pastelMixPrunedFP16 | andite/pastel-mix | Fantasy.ai | godly-devotion/MochiDiffusion | godly-devotion/MochiDiffusion | 2023 |
Orange Mix
| Google Drive Link | Original Model | License | Run on mac | Conversion Script | Year |
|---|---|---|---|---|---|
| AOM3_orangemixs | WarriorMama777/OrangeMixs | CreativeML OpenRAIL-M | godly-devotion/MochiDiffusion | godly-devotion/MochiDiffusion | 2023 |
Counterfeit
| Google Drive Link | Original Model | License | Run on mac | Conversion Script | Year |
|---|---|---|---|---|---|
| Counterfeit-V2.5 | gsdf/Counterfeit-V2.5 | - | godly-devotion/MochiDiffusion | godly-devotion/MochiDiffusion | 2023 |
anything-v4
| Google Drive Link | Original Model | License | Run on mac | Conversion Script | Year |
|---|---|---|---|---|---|
| anything-v4.5 | andite/anything-v4.0 | Fantasy.ai | godly-devotion/MochiDiffusion | godly-devotion/MochiDiffusion | 2023 |
Openjourney
| Google Drive Link | Original Model | License | Run on mac | Conversion Script | Year |
|---|---|---|---|---|---|
| Openjourney | prompthero/openjourney | - | godly-devotion/MochiDiffusion | godly-devotion/MochiDiffusion | 2023 |
dreamlike-photoreal-2
| Google Drive Link | Original Model | License | Run on mac | Conversion Script | Year |
|---|---|---|---|---|---|
| dreamlike-photoreal-2.0 | dreamlike-art/dreamlike-photoreal-2.0 | CreativeML OpenRAIL-M | godly-devotion/MochiDiffusion | godly-devotion/MochiDiffusion | 2023 |
Image Colorization
DDColor Tiny
DDColor — AI image colorization for grayscale/B&W photos using dual decoders (ICCV 2023).
| Input | Output |
|---|---|
| Download Link | Size | Input | Output | Original Project | License | Year | Sample Project | Conversion Script |
|---|---|---|---|---|---|---|---|---|
| DDColor_Tiny.mlpackage.zip | 242 MB | 512×512 RGB | AB channels (LAB) | piddnad/DDColor | Apache-2.0 | 2023 | DDColorDemo | convert_ddcolor.py |
Face Recognition
AdaFace IR-18
AdaFace — Quality-adaptive face recognition. Outputs 512-dim embedding for face verification and identification.
| Download Link | Size | Input | Output | Original Project | License | Year | Sample Project | Conversion Script |
|---|---|---|---|---|---|---|---|---|
| AdaFace_IR18.mlpackage.zip | 48 MB | Image (112×112 face) | 512-dim L2-normalized embedding | mk-minchul/AdaFace | MIT | 2022 | AdaFaceDemo | convert_adaface.py |
3D Face Pose Estimation
3DDFA_V2
3DDFA_V2 — 3D face reconstruction and head pose estimation (yaw, pitch, roll) from a single face image.
| Download Link | Size | Input | Output | Original Project | License | Year | Sample Project |
|---|---|---|---|---|---|---|---|
| 3DDFA_V2.mlpackage.zip | 6.3 MB | Image (120×120 RGB) | 62 params (12 pose + 40 shape + 10 expression) | cleardusk/3DDFA_V2 | MIT | 2020 | Face3DDemo |
Speaker Diarization
pyannote segmentation-3.0
pyannote segmentation — Speaker diarization with up to 3 simultaneous speakers. Identifies who speaks when, with overlap detection and per-speaker transcription.
| Download Link | Size | Input | Output | Original Project | License | Year | Sample Project | Conversion Script |
|---|---|---|---|---|---|---|---|---|
| SpeakerSegmentation.mlpackage.zip | 5.8 MB | 10s mono 16kHz [1,1,160000] | [1, 589, 7] speaker logits | pyannote/segmentation-3.0 | MIT | 2023 | DiarizationDemo | convert_diarization.py |
Voice Conversion
OpenVoice V2
OpenVoice — Zero-shot voice conversion. Record source and target voice, convert on-device.
| Download Link | Size | Input | Output | Original Project | License | Year | Sample Project | Conversion Script |
|---|---|---|---|---|---|---|---|---|
| OpenVoice_SpeakerEncoder.mlpackage.zip | 1.7 MB | Spectrogram [1, T, 513] | 256-dim speaker embedding | myshell-ai/OpenVoice | MIT | 2024 | OpenVoiceDemo | convert_openvoice.py |
| OpenVoice_VoiceConverter.mlpackage.zip | 64 MB | Spectrogram + speaker embeddings | Waveform audio (22050 Hz) |
Audio Source Separation
HTDemucs
Hybrid Transformer Demucs — separates music into 4 stems: drums, bass, vocals, and other instruments.
| Download Link | Size | Input | Output | Original Project | License | Year | Sample Project | Conversion Script |
|---|---|---|---|---|---|---|---|---|
| HTDemucs_SourceSeparation_F32.mlpackage.zip | 80 MB | Audio Waveform [1, 2, 343980] at 44.1kHz | 4 stems (drums, bass, other, vocals) stereo | facebookresearch/demucs | MIT | 2022 | DemucsDemo | convert_htdemucs.py |
Vision-Language
Florence-2-base
Microsoft Florence-2 — a unified vision-language model supporting image captioning, OCR, and object detection from a single model. Converted as 3 CoreML models (INT8): Vision Encoder (DaViT), Text Encoder (BART), and Decoder with autoregressive generation.
| Download Link | Size | Input | Output | Original Project | License | Year | Sample Project | Conversion Script |
|---|---|---|---|---|---|---|---|---|
| Florence2VisionEncoder / TextEncoder / Decoder | 260 MB (INT8, 3 models total) | 768x768 RGB image + task prompt | Generated text (caption, OCR, etc.) | microsoft/Florence-2-base | MIT | 2024 | Florence2Demo | convert_florence2.py |
Language Model
john-rocky/CoreML-LLM — Companion repository for running LLMs on the Apple Neural Engine. Unlike MLX Swift (GPU-only), CoreML-LLM targets ANE for ~10x lower power draw, making always-on on-device LLMs practical on iPhone. Current release v1.4.0 — Gemma 4 E2B 3-chunk decode (31.6 → 34.2 tok/s, +8.2%), chunk pipelining default ON, still-image vision encoder on ANE. All models below load via the same CoreMLLLM.load(...) Swift API and are available in-app through the Models Zoo hub.
| Model | Size | Modalities | iPhone 17 Pro decode | HuggingFace |
|---|---|---|---|---|
| Gemma 4 E2B | 3.1 GB | Text + image + audio + video | 31–34 tok/s | mlboydaisuke/gemma-4-E2B-coreml |
| Gemma 4 E4B | 5.5 GB | Text | ~14 tok/s | mlboydaisuke/gemma-4-E4B-coreml |
| Qwen3.5 2B | 2.4 GB | Text | ~17 tok/s (~200 MB RSS) | mlboydaisuke/qwen3.5-2B-CoreML |
| Qwen3.5 0.8B | 754 MB | Text | ~20 tok/s | mlboydaisuke/qwen3.5-0.8B-CoreML |
| Qwen3-VL 2B | 4.7 GB | Text + image | ~7.5 tok/s | mlboydaisuke/qwen3-vl-2b-coreml |
Gemma 4 E2B (CoreML-LLM)
Google Gemma 4 E2B (2.3B effective parameters with Per-Layer Embeddings) running fully on ANE. Multimodal: text, image (native 384x384 encoder, 196 tokens/image), audio (12-layer Conformer encoder), and video (64 tokens/frame). 2048 context length, Sliding Window Attention (28/35 layers are O(W)), PLE computed inside the graph. The default 4-chunk decode ships at 31.6 tok/s on iPhone 17 Pro; LLM_3CHUNK=1 bumps it to 34.2 tok/s by collapsing chunk 2+3.
| Download Link | Size | Input | Output | Original Project | License | Year | Sample Project | Swift Package |
|---|---|---|---|---|---|---|---|---|
| mlboydaisuke/gemma-4-E2B-coreml | 3.1 GB (INT4, 4 chunks + vision + audio + video encoders) | Text + image + audio + video (≤2048 tokens) | Generated text (streaming) | google/gemma-3n-E2B-it | Gemma ToU | 2025 | CoreMLLLMChat | CoreML-LLM |
Gemma 4 E4B
Larger text-only Gemma 4 variant — 42-layer decoder, ~4B effective parameters, 100% ANE-resident. Use when you want maximum text quality and have the storage budget. No vision / audio / video encoders.
| Download Link | Size | Input | Output | Original Project | License | Year | Sample Project | Swift Package |
|---|---|---|---|---|---|---|---|---|
| mlboydaisuke/gemma-4-E4B-coreml | 5.5 GB (INT4, 4 chunks) | Text prompt (≤2048 tokens) | Generated text (streaming) | google/gemma-3n-E4B-it | Gemma ToU | 2025 | CoreMLLLMChat | CoreML-LLM |
Qwen3.5 2B
Alibaba Qwen3.5 2B — hybrid Gated-DeltaNet SSM + attention. Shipped as 4 INT8 body chunks (6 layers each) + tail + mmap fp16 embed sidecar so a 2B-param model fits in ~200 MB phys_footprint. The 4-chunk split is required to stay ANE-resident — a 2-chunk variant at 2 GB fp16/chunk exceeds the single-mlprogram budget and falls to GPU.
| Download Link | Size | Input | Output | Original Project | License | Year | Sample Project | Swift Package |
|---|---|---|---|---|---|---|---|---|
| mlboydaisuke/qwen3.5-2B-CoreML | 2.4 GB (INT8, 4 chunks + embed) | Text prompt | Generated text (streaming) | Qwen/Qwen3.5-2B | Apache-2.0 | 2025 | CoreMLLLMChat | CoreML-LLM |
Qwen3.5 0.8B
Compact hybrid SSM+attention model, INT8 palettized — same semantic precision as fp16 (top-3 = 100% parity vs fp32 oracle), half the bundle size. Smallest and fastest option in the lineup at 754 MB / ~20 tok/s decode.
| Download Link | Size | Input | Output | Original Project | License | Year | Sample Project | Swift Package |
|---|---|---|---|---|---|---|---|---|
| mlboydaisuke/qwen3.5-0.8B-CoreML | 754 MB (INT8 palettized) | Text prompt | Generated text (streaming) | Qwen/Qwen3.5-0.8B | Apache-2.0 | 2025 | CoreMLLLMChat | CoreML-LLM |
Qwen3-VL 2B
Qwen3-VL multimodal — text + image input with DeepStack injection at L0/1/2 and interleaved mRoPE for the 196 image tokens. 28-layer GQA text backbone shipped as 6 INT8 body chunks + chunk_head + raw fp16 embed sidecar that Swift mmaps. Vision tower re-uses Qwen3-VL's native ViT.
| Download Link | Size | Input | Output | Original Project | License | Year | Sample Project | Swift Package |
|---|---|---|---|---|---|---|---|---|
| mlboydaisuke/qwen3-vl-2b-coreml | 4.7 GB (INT8, 6 body chunks + head + embed) | Text + image | Generated text (streaming) | Qwen/Qwen3-VL-2B-Instruct | Apache-2.0 | 2025 | CoreMLLLMChat | CoreML-LLM |
See CoreML-LLM for the full conversion pipeline, ANE optimization techniques (cat-trick RMSNorm, Conv2d Linear, pre-computed RoPE, stateless KV with explicit I/O), and the Swift sample app.
Zero-Shot Image Classification
SigLIP ViT-B/16
Google SigLIP — sigmoid-based contrastive image-text model for zero-shot classification. Type any labels (e.g. "cat, dog, car") and get per-label probabilities. Converted as 2 CoreML models (INT8): Image Encoder and Text Encoder.
| Download Link | Size | Input | Output | Original Project | License | Year | Sample Project | Conversion Script |
|---|---|---|---|---|---|---|---|---|
| SigLIP_ImageEncoder / TextEncoder | 386 MB (FP16, 2 models total) | 224x224 RGB image + text labels | Per-label similarity scores (softmax) | google/siglip-base-patch16-224 | Apache-2.0 | 2024 | SigLIPDemo | convert_siglip.py |
Text-to-Speech
Kokoro-82M
hexgrad/Kokoro-82M — open-weight 82M-parameter TTS by hexgrad. Style-conditioned StyleTTS2 architecture (BERT + duration predictor + iSTFTNet vocoder) producing 24kHz speech in 9 languages from per-voice style embeddings. The first CoreML port with on-device bilingual (English + Japanese) free-text input — no MLX, no MeCab, no IPADic, no Python G2P at runtime.
2 CoreML models: a flexible-length Predictor (BERT + LSTM duration head + text encoder) and 3 fixed-shape Decoder buckets (128 / 256 / 512 frames). The Swift pipeline picks the smallest bucket that fits the predicted total duration, pads input features with zeros, and trims the output audio.
| Download Link | Size | Input | Output | Original Project | License | Year | Sample Project | Conversion Script |
|---|---|---|---|---|---|---|---|---|
| Kokoro_Predictor.mlpackage.zip | 75 MB | input_ids [1, T≤256] (int32) + ref_s_style [1, 128] | duration [1, T] + d_for_align [1, 640, T] + t_en [1, 512, T] | hexgrad/Kokoro-82M | Apache-2.0 | 2025 | KokoroDemo | convert_kokoro.py |
| Kokoro_Decoder_128.mlpackage.zip | 238 MB | en_aligned [1, 640, 128] + asr_aligned [1, 512, 128] + ref_s [1, 256] | audio [1, 76800] @ 24kHz | |||||
| Kokoro_Decoder_256.mlpackage.zip | 241 MB | en_aligned [1, 640, 256] + asr_aligned [1, 512, 256] + ref_s [1, 256] | audio [1, 153600] @ 24kHz | |||||
| Kokoro_Decoder_512.mlpackage.zip | 246 MB | en_aligned [1, 640, 512] + asr_aligned [1, 512, 512] + ref_s [1, 256] | audio [1, 307200] @ 24kHz |
See sample_apps/KokoroDemo/README.md for the on-device G2P (English + Japanese), bucketed decoder strategy, and conversion details.
Anomaly Detection
EfficientAD
EfficientAD (PDN-Small) — lightweight unsupervised anomaly detection for industrial inspection. Wraps teacher, student, and autoencoder networks into a single model that outputs a per-pixel anomaly heatmap and image-level anomaly score. Pretrained on MVTec AD bottle category.
| Download Link | Size | Input | Output | Original Project | License | Year | Sample Project | Conversion Script |
|---|---|---|---|---|---|---|---|---|
| EfficientAD_Bottle.mlpackage.zip | 15 MB (FP16) | 256x256 RGB image | anomaly_map [1,1,256,256] + anomaly_score [0-1] | nelson1425/EfficientAD | MIT | 2023 | EfficientADDemo | convert_efficientad.py |
Music Transcription
Basic Pitch
spotify/basic-pitch — polyphonic Automatic Music Transcription. Converts any audio (any instrument, any voice) into MIDI notes with pitch bend detection. Just 17K parameters / 272 KB — runs in real time on iPhone with full ANE acceleration.
The first open-source iOS implementation. Loads any audio file, runs the CoreML model in 2-second sliding windows, then runs the full Python note_creation.py pipeline natively in Swift (onset inference, greedy backwards-in-time tracking, melodia trick, pitch bend extraction). Detected notes are visualized as a piano roll, exported as a Standard MIDI File, and played back through a built-in additive sine synth so you can A/B compare with the original audio.
| Download Link | Size | Input | Output | Original Project | License | Year | Sample Project |
|---|---|---|---|---|---|---|---|
| BasicPitch_nmp.mlpackage.zip | 272 KB | audio waveform [1, 43844, 1] @ 22050 Hz mono | note [1,172,88] + onset [1,172,88] + contour [1,172,264] | spotify/basic-pitch | Apache-2.0 | 2022 | BasicPitchDemo |
See sample_apps/BasicPitchDemo/README.md for the sliding-window inference, post-processing port, and iOS-specific gotchas.
Text-to-Music Generation
Stable Audio Open Small
stabilityai/stable-audio-open-small — text-to-music generation (497M params). Generates up to 11.9 seconds of stereo 44.1kHz audio from text prompts using rectified flow diffusion.
4 CoreML models: T5 text encoder, NumberEmbedder (seconds conditioning), DiT (diffusion transformer), and VAE decoder (Oobleck).
| Download Link | Size | Input | Output | Original Project | License | Year | Sample Project | Conversion Script |
|---|---|---|---|---|---|---|---|---|
| StableAudioT5Encoder.mlpackage.zip | 105 MB | input_ids [1, 64] | text_embeddings [1, 64, 768] | stabilityai/stable-audio-open-small | Stability AI Community | 2024 | StableAudioDemo | convert_stable_audio.py |
| StableAudioNumberEmbedder.mlpackage.zip | 396 KB | normalized_seconds [1] | seconds_embedding [1, 768] | |||||
| StableAudioDiT.mlpackage.zip | 326 MB | latent [1,64,256] + timestep + conditioning | velocity [1,64,256] | |||||
| StableAudioDiT_FP32.mlpackage.zip | 1.3 GB | latent [1,64,256] + timestep + conditioning | velocity [1,64,256] | |||||
| StableAudioVAEDecoder.mlpackage.zip | 149 MB | latent [1, 64, 256] | stereo audio [1, 2, 524288] at 44.1kHz |
See sample_apps/StableAudioDemo/README.md for INT8 vs FP32 DiT selection and conversion details.
Models converted by someone other than me.
Stable Diffusion
How to use in a xcode project.
Option 1,implement Vision request.
import Vision
lazy var coreMLRequest:VNCoreMLRequest = {
let model = try! VNCoreMLModel(for: modelname().model)
let request = VNCoreMLRequest(model: model, completionHandler: self.coreMLCompletionHandler)
return request
}()
let handler = VNImageRequestHandler(ciImage: ciimage,options: [:])
DispatchQueue.global(qos: .userInitiated).async {
try? handler.perform([coreMLRequest])
}
If the model has Image type output:
let result = request?.results?.first as! VNPixelBufferObservation
let uiimage = UIImage(ciImage: CIImage(cvPixelBuffer: result.pixelBuffer))
Else the model has Multiarray type output:
For visualizing multiArray as image, Mr. Hollance’s “CoreML Helpers” are very convenient. CoreML Helpers
Converting from MultiArray to Image with CoreML Helpers.
func coreMLCompletionHandler（request：VNRequest？、error：Error？）{
let = coreMLRequest.results？.first as！VNCoreMLFeatureValueObservation
let multiArray = result.featureValue.multiArrayValue
let cgimage = multiArray？.cgImage（min：-1、max：1、channel：nil）
Option 2,Use CoreGANContainer. You can use models with dragging&dropping into the container project.
Make the model lighter
You can make the model size lighter with Quantization if you want. https://coremltools.readme.io/docs/quantization
The lower the number of bits, more the chances of degrading the model accuracy. The loss in accuracy varies with the model.
import coremltools as ct
from coremltools.models.neural_network import quantization_utils
# load full precision model
model_fp32 = ct.models.MLModel('model.mlmodel')
model_fp16 = quantization_utils.quantize_weights(model_fp32, nbits=16)
# nbits can be 16(half size model), 8(1/4), 4(1/8), 2, 1
quantized sample (U2Net)
InputImage / nbits=32(original) / nbits=16 / nbits=8 / nbits=4
Thanks
Cover image was taken from Ghibli free images.
On YOLOv5 convertion, dbsystel/yolov5-coreml-tools give me the super inteligent convert script.
And all of original projects
Auther
Daisuke Majima Freelance engineer. iOS/MachineLearning/AR I can work on mobile ML projects and AR project. Feel free to contact: rockyshikoku@gmail.com