Compose Multiplatform + Native ML = Next-Gen Segmentation
Built a Subject Segmentation app using Compose Multiplatform (Kotlin) — sharing UI and core logic across both Android & iOS, while using the best native ML tools on each platform.
Features:
- Shared UI with Compose Multiplatform
- Camera permission + live camera preview
- Alignment & tilt detection before capture
- Native ML processing:
- ML Kit on Android
- Vision Framework on iOS
Where Cross-Platform Meets Computer Vision
🔗 GitHub Link: https://lnkd.in/d7YknzpE
Would love to hear your thoughts, suggestions, and improvements! 🙌
https://www.linkedin.com/posts/muhammad-waqas-4399361a3_composemultiplatform-kotlin-a[…]CoAAC-oee0BKroMb3OyHPhr44KlOzC-zLAcxPE&utm_campaign=whatsapp
s
Stefan Oltmann
11/20/2025, 6:24 AM
Sadly desktop is always left out here. 🫤
For example the only solution to face detection I found for Java is DeepJavaLibrary‘s sample.
There was kotlin-dl, but that has been abandoned two years ago.
m
M.WAQAS
11/20/2025, 6:28 AM
Thanks for sharing! Yes, desktop support is still limited in this area. We’ll review possible alternatives or desktop-friendly approaches and see what we can provide.
👍 1
s
Stefan Oltmann
11/20/2025, 6:32 AM
Sounds great. In general PyTorch, Tensorflow and ONNX Runtime are available. So there is no lack of runtimes.
The issue for me is that I don’t know how to use a ML model. You need custom code to prepare the inputs and interpret the outputs.
For face detection I use their sample because it works and they did all they necessary work for that. To be honest, I don’t understand at all what’s happening there. 😅
Stefan Oltmann
11/20/2025, 6:34 AM
For subject segmentation on Desktop I don’t know an option, so yours would be the first I think.
m
M.WAQAS
11/20/2025, 6:38 AM
Totally understand — preparing inputs/outputs for ML models can be confusing, especially without clear samples. That’s exactly why we’re planning to provide a simple end-to-end example, so you don’t have to worry about the low-level details.
If subject segmentation on desktop is still missing, we’ll aim to make ours the first easy-to-use example.