About Mushroom AI

Overview

Mushroom AI helps you identify wild mushrooms using machine learning. You can use it through the iOS app (fully on-device) or this web app (via our API). Both use the same underlying model trained on a large mushroom image dataset.

How It Works

The model is a vision-transformer trained to recognize visual features in mushroom photos. It returns confidence-ranked predictions to help you figure out what you're looking at.

iOS App: Runs entirely on your device using Apple's MLX framework. No internet required, and your photos never leave your phone.

Web App: Sends your photos to our API, which runs inference on our own hardware using PyTorch Your images are processed and immediately discarded; we don't store them.

Model Details

The model is based on DINOv3, an open-source vision transformer from Meta AI. We fine-tuned the full ViT backbone for multi-class classification across 2,514 mushroom species.

The classifier uses a linear layer on both the class token and mean-pooled patch tokens, combining global and local features. During training focal loss (Lin et al., 2017), was used, which helps with class imbalance and improves performance on rare species.

• Base architecture: DINOv3 ViT-H+/16 (Meta AI)
• Training framework: PyTorch
• iOS inference: MLX (Apple silicon, fully on-device)
• Web inference: PyTorch
• Training objective: Full fine-tuning with a linear head over class + mean patch tokens, optimized with focal loss

Training Data

The model was trained on roughly 500,000 labeled mushroom images from two major open datasets:

• iNaturalist Open Dataset - One of the largest public biodiversity datasets, with over 70 million photos of living organisms. Observations are community-contributed under various Creative Commons licenses.
• FungiTastic Dataset - Around 350,000 expert-labeled observations across 5,000 fungal species, collected over 20 years (BSD-3-Clause License; Picek et al., 2024).

These datasets provide diversity in lighting, regions, and species rarity, which helps the model generalize better.

Acknowledgments

The classification model builds on Meta's open-source DINOv3 research (© Meta Platforms, Inc., licensed under the DINOv3 License).

• Datasets: iNaturalist Open Data and FungiTastic
• Frameworks: PyTorch (training and web inference), MLX (iOS inference)

Purpose and Responsibility

Mushroom AI is for educational and informational purposes only. The model aims for accuracy, but you should always verify identifications and consult experts before handling or eating any wild mushroom.

Contact

For questions or collaborations:

support@mushroomai.app