On-Device Image Captioning

Lightweight on-device image captioning system based on MobileVLM, supporting PyTorch, ONNX, and iOS deployment.

Overview

This project implements an image captioning pipeline designed for deployment in resource-constrained environments.

• Supports multiple models: BLIP, InstructBLIP, and MobileVLM v2
• Supports PyTorch, ONNX, and Core ML deployment for MobileVLM v2

Features

• Lightweight on-device image captioning
• MobileVLM v2 support
• PyTorch / ONNX / Core ML inference
• iOS deployment support
• Edge-device optimization

Model Weights

PyTorch

The following pretrained models from Hugging Face are used in this project:

• BLIP: "Salesforce/blip-image-captioning-base"
• InstructBLIP: "Salesforce/instructblip-flan-t5-xl"
• MobileVLM: "mtgv/MobileVLM_V2-1.7B"

ONNX

Due to their large size (~6GB), ONNX model weights are not included in this repository.

Instead, you can generate them locally using the provided export script:

python src/mobilevlm/export/export_onnx/export_*.py

Core ML

Due to their large size (~6GB), Core ML model weights are not included in this repository.

Instead, you can generate them locally using the provided export script:

python src/mobilevlm/export/export_coreml/export_*.py

Running MobileVLM

https://github.com/kc-ml2/captioning_edgedevice/blob/main/document/how_to_run.md

Result

Input

sample.jpg

Question:
"What objects are visible in the scene?"

Generated Caption

"In the image, there is a living room with a fireplace, a television, a table, chairs, and a woman standing in the kitchen."

Performance and Latency

Load Latency
The model server loads the model once at startup and performs an initial warm-up pass.

VRAM Usage
GPU (NVIDIA TITAN V) memory usage is measured after the model is fully loaded and warmed up.

Inference Latency
Image captioning latency per image is measured over 500 randomly sampled images from COCO val2017.

Model Comparison

MobileVLM CPU Runtime Breakdown

Latency breakdown for a single image inference.

The current Xcode/Core ML implementation is not fully optimized.

Acknowledgement

This project is partially based on the official MobileVLM repository: https://github.com/Meituan-AutoML/MobileVLM

We adapted and modified the original implementation for:

• PyTorch runtime
• ONNX deployment
• Core ML deployment
• On-device inference optimization

License

Copyright (c) ML2. All rights reserved.

This project includes code derived from the MobileVLM repository, which is licensed under the Apache License 2.0.

The official PyTorch-based MobileVLM was modified and exported to ONNX and Core ML for on-device inference.