egotoadcitadel / palette-from-image-ai

---
## 🎨 Overview
**palette-from-image-ai** is a Windows desktop tool that reads any image — a screenshot, a photo, a game capture, a texture, a painting — and turns it into a usable color palette. Not five swatches scraped off a Wikipedia page either. The embedded vision model segments the image, clusters color by perceptual distance, weights each cluster by how much of the frame it actually owns, and emits a palette you can drop straight into a shader, a CSS file, an ASE swatch, a Godot theme, a clip-studio brush set, or a materials list.
It runs offline. It runs on one `.exe`. You extract the archive, double-click, and you're looking at a palette in under six seconds.
| Category | Details |
|---|---|
| **Product name** | palette-from-image-ai |
| **Type** | Standalone Windows desktop application (`.exe`) |
| **Core engine** | On-device vision model + perceptual color quantizer |
| **Input** | PNG, JPG, WEBP, BMP, TIFF, GIF, AVIF |
| **Output** | HEX, RGB, HSL, LAB, ASE, GPL, CSS vars, JSON, PNG swatch sheet |
| **Distribution** | GitHub landing page → download → extract → run |
| **Install footprint** | Portable. No installer, no registry writes, no runtime deps |
| **Year** | 2026 release line |
The whole point is that palette extraction stopped being a toy when the model got good at *understanding* what's in the frame. A sunset photo doesn't hand you four near-identical oranges anymore — it hands you the horizon glow, the deep water, the sand shadow, and the cloud highlight as separate, meaningful tones.
---
## 📖 What is palette-from-image-ai
A glossary, because "palette" means different things to different people and this tool means all of them at once.
| Term | Explanation |
|---|---|
| **Palette** | An ordered set of colors, usually 5–16, each with a role (dominant, accent, background, shadow, highlight). |
| **Segmentation** | The model splitting an image into regions that *belong together* — sky vs. subject vs. ground — before sampling color. |
| **Perceptual distance** | Measuring how different two colors look to a human eye, not how different their RGB numbers are. Uses CIEDE2000 in LAB space. |
| **Cluster weight** | How much of the image a color family actually occupies, used to rank palette order. |
| **Harmony mode** | Forcing the output toward a named scheme — analogous, complementary, triadic, or split-complement. |
| **Role tagging** | Auto-labeling each swatch: this is the base, this is the accent, this is the dark you'll use for outlines. |
| **Swatch sheet** | A rendered PNG of your palette with hex labels, ready to paste into a doc or Discord. |
**What you get out of it:**
- A palette that respects *regions*, not just raw pixel frequency.
- Role-tagged swatches so you know which color does which job.
- Harmony constraints when you want the output to look designed, not sampled.
- Twelve export formats so the palette drops into whatever pipeline you already have.
- The whole thing runs on your machine. No upload, no account,