The aim was to explore detail constancy in dense non photographic stylised scenes. The visual brief combined the medieval architecture of Northern Europe with the sunny lushness of a Mediterranean-style city. I wanted to feed simple layout images into an offline bespoke model, directly controlling the rendering style and lighting through crafted prompts. Additionally, I incorporated an extra scenery lora model and enhanced the detail using upscaling passes.

​

The results were partially successful. The style captured the concept's essence, especially its mood, creating attractive scenes. But on closer inspection these unedited images struggle to maintain correct object scaling in the distance, leading to illogical scenery. Due to the lack of accurate real world fantasy data, such styles may always have a small degree of design 'confusion', and will strongly benefit from more human direction.

The January 2025 video renders still struggle to stabilise and fill in all fine details - especially people. But these are a big improvement from even a few months ago.

​

Ultimately, while these over detailed images serve more as a benchmark test than accurate production art, they provide valuable insights into AI's current capabilities and limitations at the end of 2024.