README.md

---
title: Iconclass 9B
emoji: 🖼️
colorFrom: indigo
colorTo: gray
sdk: gradio
sdk_version: 6.16.0
app_file: app.py
pinned: false
short_description: Classify artworks with Iconclass codes (9B vision model)
---

# Reading pictures with Iconclass

Upload an artwork and a 9-billion-parameter vision-language model
([`davanstrien/qwen35-9b-iconclass-sft-multitask-2ep`](https://huggingface.co/davanstrien/qwen35-9b-iconclass-sft-multitask-2ep))
predicts [Iconclass](https://iconclass.org/) codes for what the image depicts. Each code is decoded to
plain English (via the `iconclass` Python package) and linked to its entry on iconclass.org.

The model was trained multi-task and the demo exposes all three modes:

1. **Classify** — the model decides which codes (and how many) apply.
2. **Cataloguing depth** — ask for *exactly N* codes (a broad 2-code record or a deep 8-code one).
3. **Catalogue completion** — paste the Iconclass codes a record *already has*; the model proposes
   additional codes only for aspects not yet covered. Most collections are partially catalogued
   (e.g. ~2 codes/object), so this matches the real cataloguing workflow.

Built with the [`gradio.Server`](https://huggingface.co/blog/introducing-gradio-server) pattern — a custom
Tufte-styled HTML frontend over a ZeroGPU inference backend. The model loads on CPU at startup and runs on
a ZeroGPU A10G inside the `@spaces.GPU` function (~7 s/inference).