Setup & Models
MLC Ollama Studio is the workspace – the images themselves are created by an image server on your network. On top of that come three optional helpers that make working with text and images easier. This page shows what you need for what.
| What for | Required? | |
|---|---|---|
| Image server | generating images | yes |
| Image analysis | turn an image into a prompt, extract a style | optional |
| Prompt improvement | turn a short idea into a detailed prompt | optional |
| Translation | write prompts in German (or 54 other languages) | optional |
1. Image server (required)
Pick the server that matches the machine that should render the images. It can be a different machine on your network than the one running the studio.
| Machine | Image server | Address in the studio |
|---|---|---|
| Mac with Apple Silicon (16 GB memory minimum, 24 GB recommended) | Z-Image (txt2zimage) – start with txt2zimage serve |
http://<machine>:11435/v1 |
| Windows PC with an NVIDIA graphics card (~25 s per image) or AMD/Intel graphics such as a Radeon 780M (~2.5 min) | MLC SD – starts the image server by itself; its Server tab shows the address to copy | http://<machine>:18081/v1 |
| Mac with Apple Silicon, alternatively | MLC SD for macOS (~52 s per image on an M2 Ultra) | http://<machine>:18081/v1 |
Connecting the studio:
- Open Settings → Server & Model Configuration → Image API (OpenAI-compatible).
- Enter the address, click Check Connection, switch on Enable and save.
- In the studio, pick
openai/z-image-turbounder Model – done. - Recommended (from 0.16.4): turn on "Generate all images via the image API (replaces Ollama)". Older recipes and history entries that still name an Ollama model such as
x/z-image-turbocan then be regenerated too – the image server always uses the model it has loaded.
From another computer, Windows must classify the image server's network as Private; on first start Windows asks once whether the server may be reachable (tick Private networks).
About Ollama: Starting with Ollama v0.32.6, Ollama temporarily no longer generates images. If you already run Ollama v0.32.5 with an image model such as
x/z-image-turbo, you can keep using it. For anything new we recommend the image server above. The optional helpers below work with any current Ollama version.
2. Optional helpers (via Ollama)
The helpers are small language and vision models that Ollama runs on your machine. Each one can be switched on separately; we tested the recommendations with the app's own prompts.
| Helper | Recommended model | Download | Alternative |
|---|---|---|---|
| Image analysis & style extraction | qwen3-vl:4b |
3.3 GB | qwen3-vl:8b (6.1 GB) recognises styles even more precisely |
| Prompt improvement | gnokit/improve-prompt |
1.6 GB | – |
| Translation | translategemma:4b |
3.3 GB | translategemma:12b (8.1 GB) for even better translations |
Install – either in a terminal:
ollama pull qwen3-vl:4b
ollama pull gnokit/improve-prompt
ollama pull translategemma:4b
… or right in the app: under Settings → Server & Model Configuration every helper page shows the recommended model with a download button.
Switch on: On the helper's page (Image Analysis, Prompt Improvement, Translation) select the model, enable it and save. Every helper can use its own server – for example image analysis on a more powerful machine.
3. Extracting a style – how it works
Found an image whose style you like? Make it reusable:
- Drag the image into the studio (or paste it with Ctrl+V).
- Optionally mark a region – otherwise the whole image counts.
- Click Extract Style, give it a name (e.g. "Van Gogh impasto") and Save & Analyze.
The studio describes only the style – technique, colours, light, brushwork – not the content. An image in the style of Van Gogh, for example, became:
oil painting, Post-Impressionist, vibrant swirling blue color scheme, impasto texture, expressive brushwork, central figure composition, dynamic swirling background
The style is saved as a brand anchor. Pick it for any prompt under Brand Anchor → Add – and your subject appears in exactly that style. Pick several brand anchors at once and the studio renders the same subject in each of those styles.
Analyze Image (instead of Extract Style) describes the whole image as a prompt – handy for rebuilding and varying an existing motif.
4. Getting started with the showcase recipes
Under Settings → Data Management → Import showcase recipes (automatic on new installations) you get six ready-made batch recipes in their own folders – the prompts behind our Showcase. Click a recipe, Configure Batch..., Run Batch.