Back to blog

tutorial

Set Up Ollama for AI Culling in SnapCuller

Alzi•

overview

What this gets you

SnapCuller's Smart Cull looks at each photo, scores it, groups bursts, and picks the keeper — so a shoot that used to take an evening can be roughed out in a few minutes. It needs a vision model to do that. The simplest free option is Ollama.

There are two ways to run it, and you only need one:

Option A — Fully offline

The model runs on your own computer. Nothing leaves the machine, no account, no internet needed after the one-time download. Best if you have a reasonably modern laptop or a GPU.

Option B — Ollama account (free hosted)

You make a free Ollama account and the model runs on Ollama's servers. Your machine stays light and it's fast even on a thin laptop. Needs internet and a free API key.

Either way, the first step is the same: install the Ollama app.

Download Ollama →

It's a normal installer for macOS, Windows and Linux. Run it once and it stays in your menu bar / system tray.

option a

Fully offline: run the model on your machine

1. Install Ollama

Download and run the installer from the link above. On macOS you'll get an app in the menu bar; on Windows it runs in the background. That's it — Ollama is now listening on http://localhost:11434, which is exactly what SnapCuller expects by default.

2. Pull a vision model

Open a terminal (Terminal on Mac, PowerShell or Command Prompt on Windows) and download one vision-capable model. We've had the best results with Gemma 3:

ollama pull gemma3

That pulls the 4-billion-parameter build (~3 GB), which is a good balance of speed and judgement on most machines. If you have a strong GPU and want sharper calls, pull a bigger one:

ollama pull gemma3:12b

The classic alternative is LLaVA, which is a bit lighter:

ollama pull llava

Rough guide: gemma3 (4b) wants ~6 GB of free RAM or VRAM; gemma3:12b wants ~10–12 GB. If a run is very slow, drop to the smaller model or use Option B.

3. Point SnapCuller at it

Open SnapCuller and go to Settings → AI. Then:

  • Choose Local (Ollama).
  • Leave the endpoint as http://localhost:11434.
  • In the model field, type the name you pulled — gemma3 or llava. Hit the refresh button next to it and SnapCuller will list what's installed.
  • Click Test connection. You want a green “Connected”.
  • Click Save & apply.

Now open a folder, turn on Auto in the Smart Cull panel to get a score on each photo as you browse, or open the Smart Cull dialog to run the whole folder at once.

option b

Ollama account: free hosted models

If your machine is light on RAM, or you just don't want a model eating your battery, run it on Ollama's servers instead. This still uses the Ollama app you installed, plus a free account.

1. Make an Ollama account

Go to ollama.com and sign up — it's free. Then sign in inside the Ollama app on your computer (menu bar icon → Sign in), which links the app to your account.

2. Create an API key

In your Ollama account settings on the website, create an API key and copy it. Keep it somewhere safe — you paste it into SnapCuller once.

3. Point SnapCuller at Ollama Cloud

In Settings → AI:

  • Choose Cloud, then Ollama Cloud.
  • Paste your API key.
  • In the model field, type a vision model that Ollama offers in the cloud — for example qwen2.5vl or llava. Check ollama.com/library for what's currently available on the cloud tier.
  • Click Test connection, then Save & apply.

The free cloud tier has usage limits. For a big wedding you might hit them mid-run — SnapCuller will pause with a clear message and you can resume later, or switch to Option A.

which one

Offline or hosted?

Pick offline (Option A) if…

  • you have ~8 GB+ of free RAM or a dedicated GPU
  • you cull on the road with no reliable internet
  • client work means nothing should leave your machine
  • you run big batches often and don't want usage caps

Pick hosted (Option B) if…

  • your laptop is light on memory or gets hot fast
  • you want it working in five minutes with no model download
  • you cull occasionally rather than every day
  • you're fine with the run needing internet

You can switch between them any time in Settings → AI — the rest of your culling workflow doesn't change.

troubleshooting

If it won't connect

“Can't reach the AI service” (local)

The Ollama app isn't running. Open it from your Applications / Start menu — you should see its icon in the menu bar or system tray. Then hit Test connection again. If you changed Ollama's port, update the endpoint in SnapCuller to match.

“Model not found”

The model name in SnapCuller doesn't match anything you've pulled. Run 

ollama list
to see exactly what's installed, and copy the name from the first column (including any :tag).

Analysis is very slow

A local model on a CPU-only machine can take several seconds per photo. Use a smaller model (gemma3 instead of gemma3:12b, or llava), or switch to Option B. For a per-photo score while you browse, you don't need to run the whole folder — just turn on Auto.

Invalid API key (cloud)

Re-copy the key from your Ollama account settings — no spaces, no trimming. Make sure you're signed in to the Ollama app as well.

recap

The short version

  1. Install the Ollama app.
  2. Offline: ollama pull gemma3, then Settings → AI → Local (Ollama) → Test → Save.
  3. Hosted: sign up at ollama.com, make an API key, then Settings → AI → Cloud → Ollama Cloud → paste key → Test → Save.
  4. Open a folder, turn on Auto or run Smart Cull.