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.

• Image Classifier

  • Efficientnetb0
  • Efficientnetv2
  • VisionTransformer
  • Conformer
  • DeiT
  • RepVGG
  • RegNet
  • MobileViTv2

• Object Detection

  • D-FINE
  • RF-DETR
  • YOLOv5s
  • YOLOv7
  • YOLOv8
  • YOLOv9
  • YOLOv10
  • YOLO11
  • YOLO26
  • YOLO-World
  • YOLOE

• Multi-Object Tracking

  • ByteTrack

• Segmentation

  • U2Net
  • IS-Net
  • RMBG1.4
  • face-parsing
  • Segformer
  • BiseNetv2
  • DNL
  • ISANet
  • FastFCN
  • GCNet
  • DANet
  • Semantic FPN
  • cloths_segmentation
  • easyportrait
  • MobileSAM
  • SAM2-Tiny
  • FastSAM

• Video Matting

  • MatAnyone

• Super Resolution

  • Real ESRGAN
  • GFPGAN
  • BSRGAN
  • A-ESRGAN
  • Beby-GAN
  • RRDN
  • Fast-SRGAN
  • ESRGAN
  • UltraSharp
  • SRGAN
  • SRResNet
  • LESRCNN
  • MMRealSR
  • DASR
  • SinSR

• Low Light Enhancement

  • StableLLVE
  • Zero-DCE
  • Retinexformer

• Image Restoration

  • MPRNet
  • MIRNetv2

• Image Generation

  • MobileStyleGAN
  • DCGAN

• Image2Image

  • Anime2Sketch
  • AnimeGAN2Face_Paint_512_v2
  • Photo2Cartoon
  • AnimeGANv2_Hayao
  • AnimeGANv2_Paprika
  • WarpGAN Caricature
  • UGATIT_selfie2anime
  • Fast-Neural-Style-Transfer
  • White_box_Cartoonization
  • FacialCartoonization

• Inpainting

  • AOT-GAN-for-Inpainting
  • Lama

• Monocular Depth Estimation

  • Depth Anything 3
  • MoGe-2
  • MiDaS

• Stable Diffusion :text2image

  • Nitro-E
  • Hyper-SD
  • stable-diffusion-v1-5
  • pastel-mix
  • Orange Mix
  • Counterfeit-V2.5
  • anything-v4.5
  • Openjourney
  • dreamlike-photoreal-2.0

• Image Colorization

  • DDColor Tiny

• Face Recognition

  • AdaFace IR-18

• 3D Face Pose Estimation

  • 3DDFA_V2

• Speaker Diarization

  • pyannote segmentation-3.0

• Voice Conversion

  • OpenVoice V2

• Text-to-Speech

  • Kokoro-82M

• Text-to-Music Generation

  • Stable Audio Open Small

• Audio Source Separation

  • HTDemucs

• Vision-Language

  • Florence-2-base

• Language Model

  • Gemma 4 E2B (text + image + audio + video)
  • Gemma 4 E4B (text)
  • Qwen3.5 2B (text)
  • Qwen3.5 0.8B (text)
  • Qwen3-VL 2B (text + image)

• Zero-Shot Image Classification

  • SigLIP ViT-B/16

• Anomaly Detection

  • EfficientAD

• Music Transcription

  • Basic Pitch

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.

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale.

Local Features Coupling Global Representations for Visual Recognition.

Data-efficient Image Transformers

Making VGG-style ConvNets Great Again

Designing Network Design Spaces

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

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

YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information. Uses PGI and GELAN architecture for efficient object detection.

YOLOv10: Real-Time End-to-End Object Detection. NMS-free architecture using consistent dual assignments — no post-processing needed.

YOLO11: Ultralytics latest YOLO with improved backbone and neck architecture. 22% fewer parameters than YOLOv8 with higher mAP.

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.

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.

YOLOE: Real-Time Open-Vocabulary Detection + Instance Segmentation. Detect and segment anything from a text query or a visual prompt (box an example object) — no fixed class list. Available in S (fast) and L (accurate). See YOLOEDemo for the region-embedding + MobileCLIP pipeline.

ByteTrack: Multi-Object Tracking by Associating Every Detection Box. Pure-Swift on-device tracker that adds persistent IDs on top of any detector above — an 8D Kalman filter plus two-stage IoU association, no appearance / ReID network.

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 - The IS-Net enhanced with our unique training scheme and proprietary dataset.

Simple and Efficient Design for Semantic Segmentation with Transformers

Bilateral Network with Guided Aggregation for Real-time Semantic Segmentation

Disentangled Non-Local Neural Networks

Interlaced Sparse Self-Attention for Semantic Segmentation

Rethinking Dilated Convolution in the Backbone for Semantic Segmentation

Non-local Networks Meet Squeeze-Excitation Networks and Beyond

Dual Attention Network for Scene Segmentation(CVPR2019)

Panoptic Feature Pyramid Networks

Code for binary segmentation of various cloths.

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

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.

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.

Fast Segment Anything — a YOLOv8-seg instance segmenter (not a SAM encoder/decoder): one forward pass segments everything and point/box prompts just select among them, the fastest SAM-family option for real-time use. FastSAM-s (light) / FastSAM-x (quality).

Note: AGPL-3.0 (Ultralytics YOLOv8), unlike the Apache-2.0 SAM family.

pq-yang/MatAnyone (CVPR 2025) — temporally consistent video matting with object-level memory propagation. From a first-frame mask it tracks and refines an alpha matte across the whole clip, holding sharp edges (hair, semitransparent regions) far better than per-frame baselines.

See sample_apps/MatAnyoneDemo/README.md for the per-frame state machine, the 5-module split, and conversion details.

Towards Real-World Blind Face Restoration with Generative Facial Prior

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

Best-Buddy GANs for Highly Detailed Image Super-Resolution

The Residual in Residual Dense Network for image super-scaling.

Fast-SRGAN.

Enhanced-SRGAN.

Pretrained: 4xESRGAN

Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network.

Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network.

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

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

Pytorch implementation of "Unsupervised Degradation Representation Learning for Blind Super-Resolution", CVPR 2021

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.

See sample_apps/SinSRDemo/README.md for the inference pipeline and conversion details.

Learning Temporal Consistency for Low Light Video Enhancement from Single Images.

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: 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

Multi-Stage Progressive Image Restoration.

Debluring

Denoising

Deraining

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

Google Drive Link	Size	Output	Original Project	Conversion Script
AnimeGAN2Face_Paint_512_v2	8.6MB	Image(Color 512 × 512)	bryandlee/animegan2-pytorch

Learning to Cartoonize Using White-box Cartoon Representations

White-box facial image cartoonizaiton

ByteDance-Seed/Depth-Anything-3 (ICLR 2026 Oral) — relative monocular depth from a single image. This Core ML port exposes only the monocular depth + confidence subgraph (camera / multi-view / sky / 3DGS branches stripped). First public Core ML conversion of DA3.

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 — unlike MiDaS-style relative depth, the depth comes out in real meters.

Left: original photo, center: metric depth (turbo colormap), right: surface normals.

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

amd/Nitro-E — AMD's 304M-parameter E-MMDiT text-to-image model. The 4-step distilled variant generates 512×512 images in ~2–3 seconds on iPhone 15+, and the full pipeline fits in ~1.04 GB after INT4 / INT8 palettization.

4-step generation on iPhone, 512×512. Prompt: "a hot air balloon in the shape of a heart, grand canyon".

3 CoreML models total:

See sample_apps/NitroEDemo/README.md for the Swift FlowMatchEulerScheduler port, tokenizer details, and iOS 18 palettization notes.

ByteDance/Hyper-SD — single-step text-to-image distilled from SD1.5 via Trajectory Segmented Consistency Distillation. ByteDance reports 2× user preference over SD-Turbo at 1 step; runs at acceptable speed and quality on iPhone 15+ via Apple's ml-stable-diffusion.

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.

See sample_apps/HyperSDDemo/README.md for the LoRA fusion, chunked-UNet palettization, and TCD scheduler details.

Pastel Mix - a stylized latent diffusion model.This model is intended to produce high-quality, highly detailed anime style with just a few prompts.

DDColor — AI image colorization for grayscale/B&W photos using dual decoders (ICCV 2023).

Input	Output

AdaFace — Quality-adaptive face recognition. Outputs 512-dim embedding for face verification and identification.

3DDFA_V2 — 3D face reconstruction and head pose estimation (yaw, pitch, roll) from a single face image.

pyannote segmentation — Speaker diarization with up to 3 simultaneous speakers. Identifies who speaks when, with overlap detection and per-speaker transcription.

OpenVoice — Zero-shot voice conversion. Record source and target voice, convert on-device.

Hybrid Transformer Demucs — separates music into 4 stems: drums, bass, vocals, and other instruments.

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.

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. All models below load via the same CoreMLLLM.load(...) Swift API and are available in-app through the Models Zoo hub.

Google Gemma 4 E2B (2.3B effective parameters) running fully on ANE. Multimodal — text, image, audio, and video input, 2048-token context. Decodes at 31–34 tok/s on iPhone 17 Pro.

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.

Alibaba Qwen3.5 2B — hybrid Gated-DeltaNet SSM + attention, INT8. Shipped as 4 chunks + an mmap fp16 embed sidecar so a 2B-param model fits in ~200 MB physical footprint and stays ANE-resident.

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.

Qwen3-VL multimodal — text + image input, re-using Qwen3-VL's native ViT vision tower. 28-layer GQA text backbone shipped as 6 INT8 chunks + an mmap fp16 embed sidecar.

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.

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.

hexgrad/Kokoro-82M — open-weight 82M-parameter StyleTTS2 TTS producing 24kHz speech in 9 languages. The first CoreML port with on-device bilingual (English + Japanese) free-text input — no MLX, no MeCab, 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.

See sample_apps/KokoroDemo/README.md for the on-device G2P (English + Japanese), bucketed decoder strategy, and conversion details.

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.

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: detected notes are shown as a piano roll, exported as a Standard MIDI File, and played back through a built-in synth for A/B comparison with the original audio.

See sample_apps/BasicPitchDemo/README.md for the sliding-window inference, post-processing port, and iOS-specific gotchas.

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).

See sample_apps/StableAudioDemo/README.md for INT8 vs FP32 DiT selection and conversion details.

apple/ml-stable-diffusion

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）

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

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

Daisuke Majima Freelance engineer. iOS/MachineLearning/AR I can work on mobile ML projects and AR project. Feel free to contact: rockyshikoku@gmail.com

GitHub Twitter Medium