Experiment · AI image generation
Stijlbeeld: AI images with a consistent style
Stijlbeeld explores a practical question in AI image generation: how do you create a series of illustrations with a recognisable visual style? A model I trained myself provides the foundation.
- Websitestijlbeeld.nl (private access)
- Statusexperimental application, private access
- At its corea custom style model, a defined visual style and controlled variations
- OutputPNG, WebP and SVG (vector)
What it does
From an idea or photograph to an illustration in one visual style
Stijlbeeld turns ideas and photographs into illustrations in a predefined style. The experiment focuses on comparing and choosing variations that work together.
Output is available as PNG and WebP, as well as SVG vector files that scale without loss of quality, from web icons to print.
Getting started
Four ways to begin
You do not need a detailed brief to get started. Stijlbeeld accepts a plain-language prompt, a URL for inspiration, an uploaded image or a photograph taken with the camera. A built-in improvement button helps refine the prompt, with advanced settings for those who want more control. The complexity stays behind straightforward controls. These screenshots come from the application itself; select an image to enlarge it.
Output
Four variations to choose from
Each request produces up to four variations in the same style. You can choose the image that best expresses your idea without having to generate another set. Each variation can be downloaded as PNG or SVG and published to the shared gallery.
Style consistency
One visual style, whatever the source
The gallery includes examples made from prompts, photographs and uploads, showing how different sources can be interpreted through the same style model.
The technology
My own model and hardware
A style model I trained myself. The visual style is learnt by the model itself. Using around two hundred reference images, I trained a custom LoRA style model on a rented NVIDIA H200 cloud GPU. This technically demanding process embeds the visual style in the model.
Generation on my own hardware. Images are generated on my own NVIDIA RTX 5090 hardware. The inputs, outputs and model remain under my control.
Why it matters
What the experiment demonstrates
Stijlbeeld explores how a custom style model and consistent settings can support a recognisable visual style. Image quality and consistency still depend on the training material, the input and human selection.
Verifiable evidence
What the application demonstrates, and its limits
- Problem
- Generated images can vary considerably in style, making it difficult to maintain a recognisable series for communications.
- Approach
- A custom style model and fixed generation settings turn prompts, URLs, uploads or photographs into controlled variations.
- Status
- An experimental application with private access. Generation runs on my own hardware.
- Demonstrable result
- Each request produces up to four variations in a consistent visual style, with PNG, WebP and SVG output.
- Limitations
- The application is not publicly accessible. Image quality and consistency depend on the training material, input and human selection.
- Last reviewed
Contact
Questions about Stijlbeeld or the technology behind it?
You’re welcome to get in touch.
Prefer to contact me directly? Call +31 6 1400 2312 or email mwd@mahieu.nl.