Field Log: AI Workflows • 7 min read
App icons define the core visual style of a brand. For MoolaMochi, we needed clean finance icons with cream backgrounds, navy outlines, and red highlights. However, some generated assets had messy borders. We used Codex and GPT Image Gen 2 to clean them up and keep them consistent.
The Problem: The Halo Effect of Naive Cleanup
Using simple background removal tools on dark layout themes is a problem. They often leave messy white borders (halos) or accidentally delete internal shapes. We started by scanning our icon assets and writing a custom script that only targets outside edges, keeping the internal details intact.

When Cleanup Fails, Regenerate
Sometimes background cleanup ruins outlines, like the paper edges of a receipt icon. When that happens, it is better to regenerate the asset. We prompted the image generator with design system constraints (thick navy outlines, rounded forms, and no backgrounds) and placed the asset on a flat chroma-key green screen. This allowed us to easily mask out the green background locally.
Consistency Comes From Auditing Sheets
To keep icons consistent across the app, place them all on a single dark-background contact sheet. This immediately highlights mismatched outline weights, weird shapes, or broken transparency edges.
The Production Workflow
- Inventory all active icon assets in your project directory first.
- Clean up transparency using edge-connected masking to protect internal colors.
- Use flat chroma-key green backgrounds for new generations to simplify local masking.
- Place all assets on a unified dark contact sheet to audit outline weights.
- Save files in optimized WebP formats and check the real app build using screenshots.
Generate like a designer, but ship like an engineer: prompt the asset, remove the matte, inspect the alpha, then verify the build.

