---
title: "Pixel LoRAs"
slug: pixel-loras
url: https://listedarticles.com/articles/pixel-loras
canonical_url: https://marceloeduardo.com/pixel-lora-training.html
content_type: blog_post
language: en
published_at: 2026-11-08T17:11:52.337Z
updated_at: 2026-11-08T17:11:52.337Z
author: "Marcelo Eduardo"
authored_by: human
publisher: "Marcelo Eduardo"
publisher_url: https://marceloeduardo.com/
topics: ["AI", "Machine Learning"]
license: all-rights-reserved
word_count: 618
reading_minutes: 3
citation: "Marcelo Eduardo, Marcelo Eduardo. \"Pixel LoRAs.\" 8 Nov 2026. https://marceloeduardo.com/pixel-lora-training.html (all-rights-reserved)"
# The full text follows. The web page shows an extract and sends readers
# to the source above; quote the citation and link the canonical URL.
---

# Pixel LoRAs

> Marcelo Eduardo explains how he trained small open image models with LoRAs to paint isometric pixel-art cities the way Codex painted his Brasília map, cutting the cost of maps like MiniDF and MiniSP from a $500-a-month Codex plan to a rented GPU or a home computer.

# Pixel LoRAs

A far cheaper way to paint pixel cities: small open image models, taught to paint the way Codex painted my map of Brasília.

The [MiniDF map](https://cadeoflavio.marceloeduardo.com/) is painted frame by frame by Codex: a 3D render of the city goes in, a hand-made isometric pixel-art tile comes out. MiniDF (6,631 frames) and its São Paulo sibling MiniSP (1,083) used up a $500-a-month Codex plan, hitting its limit twice and waiting for resets: about $0.065 and a minute of waiting per frame. Here open models learn the same move with a small LoRA (a 33 MB add-on), from pairs of 3D render and Codex painting, so any city can be painted on a rented GPU or a home computer for a fraction of that.

Every run is judged with one number, LPIPS: how far the model’s painting is from Codex’s. Lower is closer. Rounds 1–3 were judged on 32 held-out Brasília frames. Now three open models race for ten rounds each (Qwen Image 2.1, FLUX.2 klein and Qwen-Image-Edit-2511), judged on one frozen test set that also covers landmarks, São Paulo, Rio and new world landmarks.

[Current pick ↓](#pick)

[What it costs ↓](#cost)

[Training now ↓](#live)

[Test sheets ↓](#sheets)

[Open the map ↗](https://cadeoflavio.marceloeduardo.com/)

## Pick.

Leading model

Example

| Test set | Pick | Qwen Image 2.1 | 
|---|---|---|

### Why it leads

### Not settled yet

Download the LoRAs

## Cost.

| Option | All of MiniDF | vs Codex | 
|---|---|---|
| Measuring… |  |  | 

## Build.

Build your own tiny city: prompt strategies to download.

Three short guides, written for a friend who wants a map like MiniDF of his own city: the camera, the data, the prompts, how frames join without seams, and what it costs. Plain Markdown, free to read and share.

| Guide | What it covers | Size |  | 
|---|---|---|---|
| 1 · How MiniDF was made | The full pipeline: fixed isometric camera, Google 3D tiles, three zoom levels, painting each frame with Codex and stitching it seamlessly, the fixes (water, colour lock, roads, landmarks), export, hosting, cost, and the main painting prompt. | 11 KB | [Download .md](pixel-lora-training/guides/1_minidf_strategy.md) | 
| 2 · A cheap two-level version | The budget recipe: choosing heading, tilt and scale, two zoom levels, how many frames your city needs and what they cost, which APIs to enable, and the six pipeline steps up to a zoomable OpenSeadragon viewer. | 10 KB | [Download .md](pixel-lora-training/guides/2_cheap_two_level_version.md) | 
| 3 · On a Mac with the LoRAs | Painting every frame on your own Mac with the open LoRAs from this page instead of Codex: install, the exact settings, measured speed on an M5 Max, joining frames, known weaknesses, and running for days. | 7 KB | [Download .md](pixel-lora-training/guides/3_mac_m5_with_loras.md) | 

## Live.

## Models.

Test LPIPS per round · mean of the Brasília, landmark, São Paulo, Rio and world sets · lower is closer

The first rounds are training.

| Model | Round | Change | Test | Brasília | Landm. | São Paulo | Rio | World | Steps | 
|---|---|---|---|---|---|---|---|---|---|
| Loading… |  |  |  |  |  |  |  |  |  | 

Reference rows are the round 1–3 Qwen models scored on the same test set. World scores appear once the world-landmark test views are painted.

## Progress.

LPIPS per run · lower is closer to Codex

Loading…

## Rounds.

Loading…

## Sheets.

## Log.

| Run | Change | LPIPS | Copy | Edges | Steps | Result | 
|---|---|---|---|---|---|---|
| Loading… |  |  |  |  |  |  | 

- LPIPS
- Perceptual distance between the model’s painting of a held-out 3D render and the real Codex painting of the same frame, averaged over 32 fixed frames (16 at 3× detail, 16 at normal detail). The only number that decides.
- Copy guard
- Distance between the output and the 3D input. Below 0.30 the model is copying the photo-like render instead of painting, and the run is discarded.
- Edges
- How far the output’s edges are from the Codex painting’s: roads and buildings in place. A run more than 0.03 worse than the best is discarded.
- Noise floor
- The baseline run twice differs by 0.0145, so a change has to beat the best by more than that to count.
