Pixel city generation with a small model and LoRAs (marceloeduardo.com)

A new technique for generating pixel art cities has emerged, utilizing small open image models enhanced with Low-Rank Adaptation (LoRA). This method seeks to replicate the artwork of Codex, which requires significant resources—specifically, a $500 monthly subscription—to create intricate maps. The MiniDF and MiniSP projects showcased this by producing over 7,700 frames using Codex. In contrast, the new approach allows artists to generate similar visual content using only a modest 33 MB LoRA on a rented GPU or even a personal computer, drastically reducing costs while enhancing accessibility. The significance of this development lies in its potential to democratize digital art creation, making it simpler and more affordable for anyone to create pixel art landscapes of their own cities. Various open models are currently competing to produce the best results, evaluated by a metric known as LPIPS, which measures how closely a generated image aligns with Codex’s original artwork. This competition among models like Qwen Image 2.1, FLUX.2 klein, and Qwen-Image-Edit-2511 highlights advances in AI-driven art generation while offering a practical framework for users to customize their own city scapes. The availability of detailed guides serves to further empower creators in utilizing this technology.