Buddy Icon Archive Preserving the tiny pictures that made the internet personal.

With thanks

Built with a lot of help

Selected software, models, and moderation support that currently help the Buddy Icon Archive collect, enrich, test, and care for its collection.

This is a homemade archive, but it was not made alone. This page thanks the developers, maintainers, and model authors whose work the archive actually uses today. Individual icons and source collections are credited on their own archive records instead.

The register is intentionally selective. It does not list generic large-company platform infrastructure or paid proprietary services, evaluated alternatives, retired experiments, or every transitive component. Qwen is included because its openly published vision model directly shapes the archive’s descriptions and tags. OpenAI is included because offering capable image-and-text moderation free of charge is a concrete kindness to small projects like this one. Each listed project remains the work of its own maintainers and contributors, under its own terms.

Software

Independent code used by the archive

Corpus database

mysql-net/MySqlConnector

Connects the production gallery and server-side workers to their local MySQL catalog.

Image handling

SixLabors/ImageSharp

Validates images, inspects animation frames, builds fingerprints, and prepares vision and moderation inputs.

Embedded SQLite

ericsink/SQLitePCL.raw

Provides the low-level SQLite bridge and bundled native integration used by local catalog work.

Model runtime

ollama/ollama

Runs the local vision-language model used by the enrichment worker behind its Ollama-compatible interface.

Verification

Community projects that help keep it honest

Test framework

xunit/xunit

The test framework used for the archive’s policy, storage, worker, renderer, and routing checks.

Models

Models and moderation support

Archive enrichment

Qwen3-VL 30B-A3B Instruct

The locally run multimodal model behind the archive’s current qwen3-vl:30b enrichment configuration. It produces descriptions, visible text, structured visual facts, and search tags from still images and animation contact sheets.

Broad moderation

OwenElliott/image-safety-classifier-s

The pinned local ONNX classifier used to score representative animation frames and prioritize detailed safety review. Its score does not set the public verdict by itself.

Detailed moderation

OpenAI omni-moderation-latest

OpenAI makes this capable image-and-text moderation model available free of charge. That is genuinely generous: it lets a small preservation project run detailed safety review without turning moderation into a budget tradeoff. Its category flags determine the archive’s current standard-or-sensitive public verdict.