Color isn't a value. It's a story, a proportion, a feeling, and a source.
A prototype color retrieval system for visual culture — turning vague aesthetic intent into searchable, comparable, and actionable color archives.
When a designer searches for "a deep blue that's professional but not depressing," stock-photo sites give context without precision. Color pickers give precision without meaning. PaletteSeek connects the two.
Instead of asking "what color is this?", it lets you ask "what palette fits this feeling?"
| Feature | What it does |
|---|---|
| 🗣️ Intent-based Retrieval | Search with natural language — e.g., "deep blue, professional, not too warm for a thesis defense." Maps fuzzy aesthetic goals to structured artwork metadata. |
| 🎯 Color Science Matching | Uses CIELab color space and Delta-E distance to find visually (not just digitally) similar palettes, weighted by each color's proportion in the artwork. |
| 📋 Complete Color Profile | Every result is a full archive: dominant colors, hue family, cultural tags, and ready-to-use design-role suggestions (background, heading, body, accent). |
| 🛡️ Curated vs. Open | A core verified collection plus an independent user-upload track. The system grows without sacrificing the trustworthiness of the curated corpus. |
| 📊 From Search to Design | Instantly copy a full HEX set or view a visual color-network report — closing the gap between aesthetic discovery and practical production. |
Users start with text, color, or mood, apply multi-dimensional filters (tone, hue family, culture, category), and receive ranked results from a hybrid scoring engine that fuses keyword relevance + tag overlap + CIELab color distance.
Every work in the result set expands into a full palette profile with proportion bars, color-network graphs, and contextual metadata so the designer understands why the palette works, not just what the colors are.
PaletteSeek explores a classic information-retrieval challenge in a new domain:
How do you index and retrieve subjective, aesthetic content?
- Semantic gap reduction: We built structured metadata (mood, style, scene, medium) that bridges human language and visual features.
- Multi-modal hybrid scoring: Three independent signals (text relevance, color distance, tag overlap) are fused into a unified ranking.
- Information quality boundary: Two-track storage (curated vs. user-uploaded) ensures that an open system can grow without contamination.
- Color science over RGB hackery: CIELab-based matching overcomes the perceptual non-uniformity of simple RGB distance.
We documented candidate next steps — including natural-language semantic profiles, vector-based indexing, and design-tool export — in prospects.md.
git clone https://github.com/ContraMundum-code/PaletteSeek.git
cd PaletteSeek
python3 -m venv .venv
source .venv/bin/activate
pip install -r 前端/requirements.txt
streamlit run 前端/app.pyThe app will be available at http://localhost:8501.
Artwork images and metadata are sourced from Wikimedia Commons and curated by the project team. Original source links are preserved where available.
PaletteSeek is a course project developed for the Information Storage and Retrieval course at Peking University, Department of Information Management.
PaletteSeek © 2026 — Peking University, Information Management