Show HN: Valencia, Spain in Sim City/Isometric View
Learn how a diffusion model can be fine‑tuned to generate consistent isometric tiles from ordinary map data.
A diffusion model trained on a curated pixel‑art style repaints a city map tile by tile, ensuring each new tile continues the painted neighbours so the whole view stays seamless. The result looks like an isometric version of Valencia, Spain, with no obvious seams between tiles. The claim is that the model can keep the city coherent while applying the artistic style.
The pipeline starts with a hand‑picked dataset of pixel‑art‑styled images, which the model learns to imitate. During generation, the model receives one tile and the already‑painted adjacent tiles as context, then predicts the next tile so that edges line up. This iterative process forces continuity across the grid, preventing mismatched borders that would break the illusion of a single scene.
The developer of Isometric NYC used a similar approach, fine‑tuning the Qwen Image Edit model for their city. Both projects rely on diffusion‑based image editing rather than classic pixel‑art techniques, meaning the output is blurred and not composed of hard‑pixel blocks. The method is not limited to Valencia; any geospatial raster can be fed into the same pipeline with a suitable style reference.
If you want to build a 3D‑drivable map, you can adopt this tile‑by‑tile diffusion workflow: gather a style dataset, fine‑tune a diffusion model, then generate tiles sequentially while feeding neighbour context. The approach reduces manual stitching and keeps artistic consistency, but you still need to handle orientation, the default view may not align north upward without additional rotation logic.
TakeawaySequential diffusion generation with neighbour conditioning yields seamless isometric maps from ordinary raster data.