ImageDescriber

A .NET 10 CLI that uses a local Ollama vision model to generate descriptions and suggested filenames for images in bulk.

What it does

Recursively scans a directory for images (.jpg, .jpeg, .png, .gif, .bmp, .webp; TIFF is skipped and counted, because Ollama can't decode it), computes a content hash for each file, and asks a local Ollama instance to caption each unique image. Descriptions, suggested filenames, and metadata are persisted in a JSON database keyed by the content hash, so identical images at different paths share one record and re-scanning skips already-described content.

No cloud APIs are called — all inference runs locally against Ollama.

Prerequisites

• Ollama running locally or on the network (defaults to http://localhost:11434).
• A vision-capable model installed in Ollama (default llama3.2-vision):

  ollama pull llama3.2-vision
  ollama serve

• .NET 10 SDK.

Installation

git clone <repo>
cd ImageDescriber
dotnet build

Usage

Without arguments the tool opens an interactive menu. All verbs can also be invoked directly.

# Interactive menu
ImageDescriber

# Scan a directory
ImageDescriber Scan -p "C:\photos"

# Scan with a custom model and remote endpoint
ImageDescriber Scan -p "C:\photos" -m llava -e http://192.168.1.100:11434

# Search stored descriptions
ImageDescriber Search -q "dog"

# Export / import the database
ImageDescriber Export -o descriptions.csv      # or .json
ImageDescriber Import -i backup.json            # or .csv

# Print database statistics
ImageDescriber Stats

Verbs

Verb	Purpose
Menu (default)	Interactive console menu.
Scan	Hash images in a directory and describe each unique one.
Search	Keyword search across stored descriptions and file names. An entry matches when every keyword starts a word, in any order: sun finds "sunset", but cat doesn't find "located".
Configure	Edit endpoint, model, concurrency, and prompt templates.
Export	Dump the database to JSON or CSV.
Import	Merge a JSON or CSV export back into the database.
Stats	Print database statistics — total descriptions, total file size, models used, date range, duplicate count, and average description length.

Common options

Option	Long form	Effect
-p	--path	Directory to scan (Scan) or default path.
-e	--endpoint	Ollama URL. Defaults to http://localhost:11434.
-m	--model	Vision model name. Defaults to llama3.2-vision.
-q	--query	Search query (Search).
-o	--output	Export file path. The extension picks the format.
-i	--input	Import file path.

Storage

Settings and the description database are stored via ktsu.AppDataStorage (typically %APPDATA%\ktsu\ImageDescriber on Windows).

License

MIT — see LICENSE.md.