Limn

Turn rough ideas into production-ready image prompts for any T2I model, with one command.

📖 Docs · 🇺🇸 English · 🇨🇳 简体中文

Limn translates plain ideas into model-optimized prompts for text-to-image generation. Pick a target model, describe what you want, and Limn rewrites your concept using that model's preferred syntax — tag‑based for SDXL, natural prose for FLUX, cinematic language for Wan‑Image, and more.

Built for developers, designers, and anyone who needs production-quality image prompts without studying each model's prompting quirks.

Features

• One prompt, any model — FLUX, SDXL, Nano Banana, Seedream, Z‑Image, Qwen Image, and Wan‑Image. Limn rewrites your idea for whichever syntax that model expects.
• Transform or generate — Optimize prompts only, or go all the way with Replicate-backed image generation. Get back a WebP file with full cost and timing analytics.
• Per-model prompt intelligence — Every model profile encodes real prompting rules: tag‑based for SDXL, prose for FLUX, reasoning‑first structure for Nano Banana. No guesswork.
• Cost analytics built-in — Token counts, generation time, and OpenRouter plus Replicate costs displayed after every run. JSON output captures every number for automation.
• Smart API key management — Keys resolve from environment variables or OS keychain. No plaintext dotfiles. Secret values are redacted by default.
• Programmatic API — Call limn(), limnGenerate(), or instantiate the Limn class with keys passed directly. TypeScript‑first with full type coverage.
• Inline image preview — Terminals that support it render the generated image inside the console after generation.
• Model catalog API — getSupportedModelCatalog() exposes all model metadata. Drive downstream selectors and validators from Limn's canonical model definitions.

Quick Start

Install and generate your first image:

npm i -g @telepat/limn
limn settings set openrouterApiKey sk-or-...
limn settings set replicateApiKey r8_...
limn -m flux "a cat in space" --generate

Expected outcome:

• Limn transforms your prompt into FLUX-optimized syntax via OpenRouter.
• Replicate generates and saves a timestamped .webp to the current directory.
• Cost and timing analytics are displayed after every run.

Requirements

• Node.js 20+
• npm
• OpenRouter API key
• Replicate API token (for generation)

How It Works

Limn loads a per-model profile that encodes the preferred prompt syntax for that T2I model. It sends your raw prompt — plus the profile's system instructions — to an LLM via OpenRouter. The LLM returns a model-optimized prompt. Optionally, Limn sends that prompt to Replicate for image generation.

Core commands:

# Transform only
limn -m sdxl "a cat in space"

# Transform and generate
limn -m flux "a cat in space" --generate

# JSON output
limn -m flux "a cat in space" --generate --json

# Custom aspect ratio
limn -m flux "a cat in space" --generate --aspect-ratio 16:9

# Override Replicate model
limn -m flux "a cat in space" --generate --replicate-model black-forest-labs/flux-2-pro

# List model options (including supported aspect ratios)
limn -m seedream "a cat in space" --list-options

Aspect Ratio Support

All models support 1:1, 16:9, and 9:16 universally. Each model also supports additional ratios depending on its capabilities — Limn validates the chosen ratio before sending it to Replicate, so you get a clear error immediately if the ratio is unsupported by that model.

Use --list-options on any model to see its full set of supported aspect ratios.

Per-model prompting strategies are documented in the Prompting Guide.

Model Comparison

The same prompt — "an owl librarian at a wooden desk in a grand victorian library, reading a glowing book by candlelight, dust motes floating in sunbeams" — transformed through each model's unique syntax and generated via Replicate:

Model	Transformed Prompt	Generated Image
Flux (schnell)	An owl librarian sits at a wooden desk in a grand Victorian library, absorbed in reading a glowing book by candlelight. Dust motes float in the sunbeams that filter through the library…
SDXL	masterpiece, best quality, highly detailed, 8k uhd, owl librarian, wooden desk, grand victorian library, reading, glowing book, candlelight, dust motes, sunbeams Neg: low quality, blurry, distorted, bad anatomy
Nano Banana	An owl librarian sits at a wooden desk in a grand Victorian library, reading from a glowing book illuminated by candlelight. Dust motes float through the air, visible in the sunbeams streaming into the room…
Seedream	An owl librarian sits at a wooden desk in a grand Victorian library, reading a glowing book by candlelight. Dust motes float through the air, illuminated by sunbeams streaming in.
Z‑Image	Owl librarian sitting at a wooden desk in a grand Victorian library, reading a glowing book by candlelight. Dust motes float in sunbeams streaming through the window.
Qwen Image	An owl librarian sits at a wooden desk in a grand Victorian library, reading a glowing book by candlelight. Dust motes float in the sunbeams. Ultra HD, 4K, cinematic composition.
Wan Image	A grand Victorian library setting, where an owl librarian is perched at a wooden desk. The scene focuses on the owl as it reads a glowing book, illuminated by the warm glow of candlelight. Sunbeams filter through the space…

Each model applies its own syntax — comma‑separated keyword tags for SDXL, front‑loaded natural prose for FLUX, cinematic scene language for Wan Image, and more. Limn handles all of this in a single command.

Using With AI Agents

Limn is designed for agentic and automated workflows:

• Machine-readable output — --json returns structured JSON for both transform-only and generate runs. Every field is typed and predictable.
• Programmatic API — The library exports limn() and limnGenerate() for functional use, plus the Limn class for object‑oriented workflows with injected API keys. No interactive prompts required.
• CI‑ready — Pass keys via TELEPAT_OPENROUTER_KEY and TELEPAT_REPLICATE_TOKEN environment variables, or through the Limn constructor. No keychain or interactive settings command needed.
• Model catalog API — getSupportedModelCatalog() exposes canonical model metadata for downstream tooling and runtime validation.

Security and Trust

• API keys are stored in the OS keychain by default via limn settings.
• In CI or containerized environments, use TELEPAT_OPENROUTER_KEY and TELEPAT_REPLICATE_TOKEN environment variables.
• Set TELEPAT_DISABLE_KEYTAR=true when keychain access is unavailable.
• limn settings list redacts secret values as ***configured***.
• Generated images are model-produced output. Review images before publishing.

To report a security issue, see the security policy or open a private report through the repository security flow.

Documentation and Support

• Repository
• Prompting Guide
• npm package
• Issues

Contributing

Contributions are welcome. Open an issue or pull request on GitHub.

License

MIT. See LICENSE.