Local AI Image Recognition: How Talyx Sees Your Photos

Talyx has always been great at classifying documents — PDFs, contracts, spreadsheets. Drop a folder, and the AI reads the content to sort and rename everything. But images? Videos? Those used to get generic names like photo_20260330.jpg. Not anymore.

The Problem: Text AI Can't Read Pixels

Text-based AI models are brilliant at reading documents — they understand that a file is an invoice or a contract based on its content. But images don't have text content (unless they're scanned documents). A photo of a cat is just a pile of pixels. Without actually seeing the image, a text model can't tell a cat from a car.

The Solution: SmolVLM — Local Vision AI

Talyx now includes SmolVLM, a vision-language model that runs entirely on your device via ONNX Runtime. When you sort a folder containing photos, SmolVLM analyzes what's in each image and returns a descriptive classification — all without sending a single byte to the cloud.

The results speak for themselves:

• image-5.jpg → white-persian-cat-sofa.jpg
• image-6.jpg → retro-orange-muscle-car-showroom.jpg
• image-9.jpg → taj-mahal-front-view.jpg
• video-1.mp4 → golden-retriever-playing-park.mp4

How It Works

The sort flow uses a multi-model pipeline, with each model handling what it does best:

1. Documents — Qwen2.5 reads text content and classifies into semantic subcategories. Runs locally via node-llama-cpp.
2. Images & Videos — SmolVLM analyzes visual content to identify objects, scenes, and context. Runs locally via ONNX Runtime.
3. Scanned documents — Tesseract.js performs OCR, then GLiNER extracts named entities (names, dates, organizations). Both run locally.
4. Fallback — Anything the AI can't classify gets sorted by file type and metadata.

No Credits. No Cloud. No Compromise.

Unlike cloud-based vision APIs that charge per image, Talyx's vision AI is included with the app. No credits to buy, no usage limits, no monthly fees. SmolVLM runs on your device — classify as many images as you want.

Talyx is $29 one-time. That includes text classification, vision AI, OCR, entity extraction, and every future model update. No subscription, ever.

Privacy

Your photos never leave your device. SmolVLM processes images in memory on your machine and the results stay local. No thumbnails are sent to any server. No cloud APIs are involved. This is especially important for personal photos, medical images, or any visual content you wouldn't want processed by a third party.

Three Presets for Every Machine

Vision AI is more demanding than text classification. Talyx includes three presets so it works well on any hardware:

• Light — Smaller models, lower memory usage. Works on older machines with 8GB RAM.
• Balanced — Good quality and reasonable speed. The default for most users.
• Max — Best classification quality. Ideal for machines with dedicated GPUs or Apple Silicon with 16GB+ RAM.

What's Next

Local vision AI is available now. Update to the latest version of Talyx and try sorting a folder with photos. You'll see the difference immediately — every image gets a descriptive name and lands in the right category, all processed 100% on your device.