AI-Powered Image Search for the TYPO3 Media Module

By @obartsch

What is your idea about?

Finding a specific image in a large TYPO3 media library today usually means already knowing (or guessing) its filename. Editors can’t search “car” and get every image that actually shows a car, because TYPO3 has no concept of what an image depicts, only what it’s called.

This isn’t a new problem in the wider CMS world, other platforms have already solved it: Drupal CMS ships AI-assisted media tagging and discovery out of the box, and Adobe Experience Manager has offered AI-based smart asset tagging for years. But every one of these implementations is cloud-only, built on a SaaS AI provider with no self-hosted option.

TYPO3 has a real opportunity here: build the one CMS-native version of this feature that takes self-hosting and data sovereignty seriously by design, not as an afterthought, which is exactly the kind of positioning TYPO3 already claims for itself.

This idea adds AI-based image understanding to the Media / File module: images get automatically analyzed by a vision model, and editors can search for image content, not just filenames. This is designed to work with self-hosted, local AI models wherever possible, so no image data needs to leave the server, with cloud providers as an optional, admin-configured alternative for those who prefer it.

It will be built as a standalone TYPO3 extension, compatible with TYPO3 13 LTS and later, so existing installations can use it right away.

What must be achieve by 30th of November 2026?

  • A background indexing pipeline that runs images through a configurable AI vision provider (local-first, cloud optional) and stores the resulting tags/description as file metadata

  • New uploads get indexed automatically; existing libraries can be (re-)indexed via a Scheduler task

  • Search integration in the Filelist/Media module so editors can find images by content, not just filename

  • An admin-configurable two-tier model: a lightweight tagging/keyword-search tier that works wherever AI is configured, plus an optional, opt-in semantic-search tier (e.g. “car” also matches “vehicle”) for installations willing to run a bit more infrastructure

  • A safe default: if no AI provider is configured at all, nothing changes and search keeps working exactly as it does in TYPO3 today

  • Clear documentation on which local models/setups are realistic on typical hosting versus what needs dedicated infrastructure

What is the potential impact of your idea for the overall goal?

  • Closes a real competitive gap. Editors coming from Drupal or Adobe-based platforms already expect to search media by content. TYPO3 currently can’t do this at all.

  • A genuine market differentiator, not just a catch-up feature. Every comparable implementation we found (Drupal CMS, Adobe Experience Manager, mainstream DAM tools) is cloud-only. Nobody in this space ships a self-hosted-first version. TYPO3 doing this properly turns “we’re behind” into “we did it the way it should have been done.”

  • Privacy by design. No image data has to leave the server, a real differentiator for TYPO3’s public-sector and privacy-conscious user base.

  • Scales to what people can actually run. Small installations get useful tagging; larger or self-hosted setups get full semantic search. Nobody is locked out, nobody is forced into infrastructure they don’t want.

  • Reusable groundwork. The AI-provider integration and indexing pipeline are generic enough that other extensions could build on the same foundation later.

How does your Idea align with the TYPO3 Association Strategy

Improves developer and editor experience by solving a real, everyday content-management pain point. Supports privacy and digital sovereignty, both core TYPO3 values, by prioritizing local/self-hosted AI over cloud dependency. Directly strengthens TYPO3’s competitive position: rather than catching up to a feature Drupal and Adobe already ship, this makes TYPO3 the one major CMS offering it in a self-hosted, sovereignty-respecting way, turning a feature gap into a differentiator.

Budget for this idea?
10.000 Euro

My Name
Oli Bartsch

1 Like

Could that also be used to “propose” alternative texts for accessibility and SEO?

Do you have any specific AI vision providers in mind for the first version of the extension?

Yes, technically it’s possible, since the extension will send images through a vision model. The generated description could just as easily seed an alt-text or SEO suggestion. That said, I’d like to keep this idea’s scope focused on generating the data needed for content search specifically. Alt-text generation is already served by other TYPO3 extensions, so building it again here would in my opinion be duplicating effort rather than closing a gap.

For local/self-hosted: Ollama, since it’s basically the most widely used way to self-host vision models today (LLaVA, Moondream, etc.) and has a simple HTTP API. For cloud, I think an OpenAI-compatible vision API might be used. However, the actual provider used is up to the administrator, since the extension doesn’t hardcode any provider and just consumes the provider’s API.

1 Like