Keeping AI Style Consistent Across 441 Images: The Isometric London Map Project
Mastbubbles · reddit · 2026-08-06
The creator used an AI image model to restyle 441 Google aerial tiles into a single pannable isometric map of London. Since getting an AI to generate multiple images with exactly the same style is notoriously difficult, the seams between tiles were highly visible.
The Engineering Fix:
- Generate in an outward spiral from the center, passing already-finished neighboring tiles as reference images to ensure local continuity.
- Use a single, fixed "anchor" tile as a permanent reference to prevent style drift across 441 sequential steps.
- Automated QA: Quality is measured numerically by correlating the output with the source. A correlation below 0.15 means the model hallucinated a fake London, triggering an auto-reroll (normal correlation is around 0.85).
This technique is applicable to movies, campaigns, and map-making. An interactive version and full tutorial are provided.
Related event: AI Used to Stitch 441 Images into London Map(2 posts)→
More from Multimodal
- Community Calls for Quantized MiniMax H3 for Apple Silicon — sockenull · 2026-08-06
- Hailuo H3 Meets Adobe AE: New Plugin Enables In-App Video Generation and Object Replacement — aziz4ai · 2026-08-06
- Developer Builds Local Multimodal AI to Generate Personalized Audiovisual Works — Merzmensch · 2026-08-06
- Video Generation Showdown: Seedance 2.5 vs. Minimax H3 — LudovicCreator · 2026-08-06
- Turning a Single Product Image into an AI Commercial: Full Workflow — Gullible-Goose-1992 · 2026-08-06
- 'I don't want to play with you anymore': Hilarious Text-to-Video Fail — Superb-Painter3302 · 2026-08-06