Turn Images Into Reusable AI Style Skills.
Download StyleDNA · Installation and release details · Report an issue
| Platform | Installer |
|---|---|
| Windows x64 | Setup EXE · MSI |
| macOS Apple Silicon | DMG · App ZIP |
| Linux x64 | AppImage · DEB · RPM |
Installers are unsigned. Trusted Windows signing and Apple notarization require publisher credentials that are not available yet. See the platform verification record and repository build record for details. Full macOS/Linux desktop UI behavior remains unverified.
Prefer to stay in Claude Code, Cursor, Copilot or another agent? StyleDNA is also available as ready-to-install agent skills on Agensi:
- StyleDNA: Image to Style Skill turns reference images into a reusable style
SKILL.md. - StyleDNA Pro adds ready-to-paste prompts for Midjourney, Stable Diffusion, Flux, GPT Image, Gemini and Ideogram, plus a checker that scores your generations against the style.
The desktop app stays free and open source.
StyleDNA is a local-first desktop workspace for collecting visual references, analyzing their style, and generating reusable SKILL.md files. It pairs a Tauri 2 / React interface with a Python image-processing service and a versioned SQLite library.
Its output is reusable SKILL.md files and style profiles. The workflow is upload references → analyze style → review and edit → save or export skills. See product scope.
- Import PNG, JPEG, and WebP references with content validation, normalized copies, and thumbnails.
- Measure real palettes, luminance, saturation, contrast, edges, and texture without an AI model.
- Optionally interpret visual traits with an installed local Ollama vision model. Results are validated suggestions; failures never fall back to invented output.
- Edit the palette, visual traits, positive rules, negative constraints, and subject exclusions.
- Save, search, categorize, tag, favorite, duplicate, delete, and restore saved skill revisions.
- Preview/copy deterministic
SKILL.mdcontent; export Markdown, JSON, or portable ZIP packages; import packages under new identities. - Dark/light themes, drag-and-drop import, keyboard navigation, progress/cancellation, and unsaved-change navigation protection.
- A comfortable, spacious studio with adjustable reference/editor panels, a collapsible Markdown preview, grouped sidebar navigation, and expandable advanced controls.
The roadmap focuses on style extraction, multi-reference analysis, skill authoring, compatibility, and desktop reliability. In-app image synthesis and generated-image history are outside the product scope. Exported style instructions are guidance for compatible AI tools and cannot guarantee exact visual reproduction.
Development requires Node.js 24 LTS, pnpm 12.21.0, Python 3.12, and uv. Desktop builds also require Rust 1.90+ and the Tauri platform prerequisites.
corepack pnpm install --frozen-lockfile
uv sync --locked
corepack pnpm devOpen http://127.0.0.1:1420. This starts the real backend and frontend together, stores development data in .styledna-dev/, and stops the backend when the launcher exits. Running dev:web alone only starts Vite; it does not start a backend.
In Windows PowerShell, use corepack.cmd in place of corepack if script execution policy blocks the PowerShell shim. No execution-policy change is needed.
For the desktop application:
corepack pnpm build:sidecar
corepack pnpm desktopThe sidecar packages Python and its runtime libraries. Installed-app users do not need Python, Node, or uv. Development does. Rebuild the sidecar after Python changes when testing the desktop host.
Actual Windows desktop workspace using the included, manually authored example reference.
Create a new skill and import examples/amber-study.png. Choose Local measurements, analyze, apply the result, and add your own visual traits. Save, then export a package.
Alternatively, import examples/amber-study.zip from the library. This example is a synthetic, manually authored illustration and profile licensed under MIT; it is not presented as AI output. Nothing is auto-added to your library.
Install Ollama and a compatible vision model separately. For example, ollama pull gemma3:4b installs a vision-capable model (about 3.4 GB; see its model page and license). Start Ollama, open Settings & Providers, save/check its loopback address, and select a model reporting vision support. In the studio choose Ollama vision and start analysis. Model discovery never downloads or runs a model automatically. Cloud-routed Ollama models are excluded.
Local measurements extract numeric evidence. They intentionally leave semantic style fields unassessed. Human edits remain useful with no model installed. Ollama interpretations require review and are not calibrated confidence scores. Multiple references currently have equal influence.
uv run pytest -q
uv run ruff check backend scripts
uv run ruff format --check backend scripts
corepack pnpm test
corepack pnpm build
corepack pnpm format:check
corepack pnpm exec playwright install chromium
corepack pnpm test:e2eNative checks require the sidecar first:
corepack pnpm build:sidecar
cargo check --manifest-path src-tauri/Cargo.toml
cargo test --manifest-path src-tauri/Cargo.toml
corepack pnpm build:desktopSee verification status for commands actually run on the development host. Workflow configuration is not proof of tested platform support. Cross-platform packaging is configured, with platform results documented separately. No releases are published automatically.
There is no telemetry, account requirement, or automatic image upload. Measurement analysis stays inside the application. Ollama analysis sends bounded reference previews to your selected loopback service only when you start it. Exports contain metadata-stripped reference copies; your originals are never modified.
The desktop app stores its database and assets under the OS application-data directory for com.styledna.desktop. Settings displays the exact path. Development uses .styledna-dev/ unless STYLEDNA_DATA_DIR is set. Portable packages are the supported per-skill backup mechanism. Quit before backing up an entire data directory; include SQLite WAL files if present.
The authenticated local service binds an ephemeral loopback port. The desktop host owns its process. Credentials are per-launch and are not persisted. The current app has no cloud-provider API keys. See SECURITY.md and architecture.
20 references per skill; 20 MiB encoded input and normalized reference limit; 40 megapixels decoded per image; 200 MiB combined image upload; 250 MiB expanded package; 5 GiB managed-asset storage. Saving also enforces a package budget with metadata overhead so a skill remains portable. Animated images are rejected. Inference previews are bounded to 2,048 pixels and thumbnails to 320 pixels. Limits live in backend/styledna/config.py.
SKILL.md uses a conservative name/description frontmatter and relative reference links. The package is portable data, not an executable plugin. Agent-specific installation/loading and generator compatibility have not been certified; review the target agent's current skill conventions before installation. Markdown/JSON exports alone do not contain the reference assets; use ZIP for a self-contained package.
MIT licensed. Contributions welcome: CONTRIBUTING.md.

