Does it use OpenAI or Gemini?

No. AI inference runs locally in the browser using a compact multilingual model and ONNX Runtime Web. Generation does not use a remote AI inference API.

Do I need an account?

No account is required to generate and export palettes.

Does it work with Russian prompts?

Yes. The model is multilingual and handles Russian and English prompts.

How does text-to-palette work?

Your text is embedded into a semantic vector using a local multilingual model. That vector is matched against a set of semantic color anchors, which maps it to an OKLCH base color. The palette is then generated from that base color using a perceptual color harmony.

Why does the first generation take longer?

Model assets (the ONNX model and tokenizer data) are downloaded on first use and cached by the browser. Subsequent generations are faster.

Local AI · No external API

AI Color Palette Generator

Type a scene, mood, or idea. A local in-browser AI model maps your text to a starting OKLCH color palette. Works in English and Russian.

Or try a prompt idea:

Example prompts

winter forest purple cave deep sea horror neon cyberpunk rain cozy autumn cafe desert ruins at dusk

How it works

Describe

Type any scene, mood, or idea in English or Russian.

Embed → Match

A local multilingual model converts your text to a semantic vector and matches it to an OKLCH color anchor.

Generate palette

The matched anchor becomes the base color. Shadow, highlight, and accent are derived by OKLCH perceptual harmony.

Technical details

AI inference runs locally in the browser using a compact multilingual text embedding model and ONNX Runtime Web. Required model assets are downloaded on first use and cached by the browser. Generation does not use a remote AI inference API — no OpenAI, Gemini, or Anthropic.

The semantic matching layer uses a set of handcrafted color anchors that map semantic concepts (cold, warm, dark, vivid, organic, etc.) to OKLCH color targets. This gives predictable, art-direction-friendly results rather than statistically arbitrary colors.

All generated colors are fitted to the sRGB gamut while preserving hue and perceptual lightness. The palette uses OKLCH Shadow, Base, Highlight, and Accent roles.

Multilingual E5 Small (ONNX)OKLCH Color SpaceSemantic Anchor MatchingsRGB Gamut Fitting

What it's useful for

Pixel art

Get a starting palette for a scene before hand-tweaking individual colors.

Illustration

Explore color mood quickly without manually navigating color wheels.

Game jams

Generate themed palettes fast when you need something coherent under time pressure.

UI/web design

Start with a semantically grounded palette, then refine individual tokens.

Limitations

• The model interprets semantic meaning, not physical color names. "Bright red" will produce a red-range result, but exact output depends on semantic similarity to trained anchors.
• Generated palettes are starting points. Pixel artists typically adjust colors manually after generation.
• First generation requires downloading model assets (~20–30 MB). An active internet connection is needed on first use.
• Generation on low-end mobile devices may be slower than on desktop.

FAQ

Does it use OpenAI or Gemini?: No. AI inference runs locally in the browser using a compact multilingual model and ONNX Runtime Web. Generation does not use a remote AI inference API.
Do I need an account?: No account is required to generate and export palettes.
Does it work with Russian prompts?: Yes. The model is multilingual and handles Russian and English prompts.
How does text-to-palette work?: Your text is embedded into a semantic vector using a local multilingual model. That vector is matched against a set of semantic color anchors, which maps it to an OKLCH base color. The palette is then generated from that base color using a perceptual color harmony.
Why does the first generation take longer?: Model assets (the ONNX model and tokenizer data) are downloaded on first use and cached by the browser. Subsequent generations are faster.