ktoyame/tiny-london-lora

tiny.london LoRA

This LoRA redraws aerial 3D renders of London as isometric pixel art. It drew the whole map at tiny.london, which covers every street in Greater London.

Built with Qwen. The LoRA is fine-tuned from Qwen-Image-2.1-Turbo and released under the Qwen Research License, so you may use it for non-commercial research and evaluation only. See LICENSE and NOTICE.

Examples

Each pair shows the 3D render on the left and the LoRA's drawing on the right.

Imagery in the left halves © 2026 Google.

How to use it

The weights are tiny-london-v3.safetensors (fp16, 159 MB). The LoRA expects the same input it was trained on.

• Input: a 1024 x 1024 px oblique aerial render, like the left halves above. We used Google Photorealistic 3D Tiles.
• Frame: draw a 2 px pure red (255, 0, 0) border round the edge of the input image. Every training input had one.
• Prompt: Fill in the outlined section with the missing pixels corresponding to the <isometric london pixel art> style, removing the border and exactly following the shape/style/structure of the surrounding image (if present).
• Sampling: 8 steps, guidance 1.0 (no CFG), LoRA weight 1.0.
• Base weights: qwen_image_2.1_turbo_int8_convrot.safetensors from Comfy-Org/Qwen-Image-2.1, as in training.

We trained and ran the LoRA with ostris/ai-toolkit (model architecture qwen_image_2). We have not tested it in other tools.

Some training inputs were part render and part finished pixel art, with the red frame round the part to draw. So the LoRA can also extend a finished drawing into a new area.

Drawing a whole map

The map at tiny.london was not drawn one tile at a time. It was drawn as one shared latent canvas: at every step, overlapping 1024 px windows (768 px apart) are denoised and their predictions are averaged into the canvas. This leaves no seams between windows. The sampler code is not part of this repository.

Training

• Tool: ai-toolkit, LoRA rank 32 (alpha 32).
• Schedule: 8,000 steps, batch size 1, AdamW 8-bit, learning rate 1e-4, bf16, flow matching with shifted timesteps.
• Base: Qwen-Image-2.1-Turbo with int8 (convrot) weights. The turbo training adapter was on during training only.
• Hardware: one H100, about 4 hours.
• Data: 6,326 examples at 1,024 px and 512 px, made from about 2,500 frames across Greater London. Inputs are renders of Google Photorealistic 3D Tiles. Targets are the same frames redrawn by Gemini 3 Pro Image, then filtered automatically to drop copies of the photo, invented water and invented buildings. Every example has the same caption, the prompt above.

Limitations

• On empty water the LoRA sometimes draws London landmarks that are not there: Big Ben, the Houses of Parliament or Tower Bridge in the middle of a river or a reservoir. On tiny.london an automatic check found them and they were painted out, apart from one kept as an Easter egg.

• It paints some fields, sports pitches and glass roofs in the colour of water.
• It was trained on London only. We have not tested it on other cities.
• It draws no people or vehicles. That was a choice made for the training targets.

Credits

• The approach and the outline prompt come from isometric-nyc by cannoneyed (MIT licence).
• Renders: Google Photorealistic 3D Tiles. Imagery © 2026 Google.
• Base model: Qwen-Image-2.1-Turbo by the Qwen team. Built with Qwen.