Isometric SF maps Bay Area in pixel art

Isometric SF is a masterclass in model cascading, proving that solo developers can tackle massive generative datasets by distilling frontier outputs into nimble 4B models. Beyond the SimCity nostalgia, the technical blueprint shows how smart data preprocessing and architectural distillation trump raw compute brute force.

• –Multi-tier distillation cut costs by ~75%: Nano Banana 2 generated 1,460 seed pairs to train a Qwen-Image-Edit 2511 LoRA, which then distilled into a 4-step FLUX.2 Klein 4B LoRA running in 1.1s per tile on H100s.
• –Hemming replaced shared latents: Instead of MultiDiffusion or latent canvas pinning that introduce tile dependencies, independent tile generation with a 64px boundary fade enabled embarrassingly parallel execution and single-tile fixes.
• –Vector-raster hybrid optimization: Using OpenStreetMap data to overlay water skipped 16% of tile painting, while custom gradient detectors smoothed color mismatches across Google's multi-season aerial imagery.
• –Coding agents orchestrated the pipeline: The entire capture, training, and eval infrastructure was coordinated via Claude Code over a Mac mini and rented cloud H100s (119 production hours).