Lumihoo AI Web App

Lumihoo is a polished image generation interface for OpenHoo. It lets users turn prompts into luminous generated images through an OpenAI-compatible SGLang image endpoint.

Features

• Prompt-driven image generation with one to four outputs.
• Preset controls for turbo, balanced, and quality generation modes.
• Configurable style presets that are injected into upstream prompts.
• 1024 and 2048 square image sizes.
• Optional deterministic seed input.
• Result gallery, lightbox preview, image download, and share actions.
• Generated image storage in MinIO with metadata in Postgres.
• Caddy-hosted image URLs for stable sharing.
• Dark, responsive Next.js interface with production analytics enabled on Vercel.
• Standalone Docker build for self-hosted deployment.

Tech Stack

• Next.js 16 and React 19
• TypeScript
• Tailwind CSS 4
• SGLang-compatible image generation API
• Drizzle ORM
• Postgres
• MinIO
• Caddy
• pnpm

Requirements

• Node.js 24 or newer
• pnpm 11
• Postgres
• MinIO
• An OpenAI-compatible SGLang image generation endpoint

The app defaults to http://localhost:30010/v1 and model ideogram-ai/ideogram-4-nf4.

Configuration

Set these environment variables as needed:

Variable	Default	Description
DATABASE_URL	unset	Postgres connection string used by Drizzle.
MINIO_ENDPOINT	localhost	MinIO endpoint hostname or URL.
MINIO_PORT	9000	MinIO API port.
MINIO_USE_SSL	false	Whether the app connects to MinIO over HTTPS.
MINIO_ACCESS_KEY	unset	MinIO access key.
MINIO_SECRET_KEY	unset	MinIO secret key.
MINIO_BUCKET	lumihoo-images	Bucket used for generated images.
MINIO_REGION	us-east-1	Region used when creating the bucket.
PUBLIC_IMAGE_BASE_URL	/images	Public URL prefix for stored image objects.
SGLANG_BASE_URL	http://localhost:30010/v1	Base URL for the SGLang OpenAI-compatible API. /v1 is appended automatically if omitted.
SGLANG_API_KEY	EMPTY	Bearer token sent to the upstream endpoint.
SGLANG_IMAGE_MODEL	ideogram-ai/ideogram-4-nf4	Image model name passed to the upstream API.
SGLANG_TIMEOUT_MS	110000	Request timeout in milliseconds. Clamped between 1000 and 300000.
IDEOGRAM_PRESET	V4_DEFAULT_20	Default preset. Supported values: V4_TURBO_12, V4_DEFAULT_20, V4_QUALITY_48.
IDEOGRAM_SIZE	1024x1024	Default size. Supported values: 1024x1024, 2048x2048.
IDEOGRAM_JSON_PROMPT	auto	Whether to wrap text prompts as Ideogram 4 structured JSON captions. Defaults to enabled when SGLANG_IMAGE_MODEL contains ideogram and 4. Set to false for non-Ideogram-compatible upstreams.
IDEOGRAM_SEED	unset	Optional non-negative integer seed.
LUMIHOO_STYLE_PRESET	natural	Default style preset for the legacy profile. Built-in values are natural, photoreal, cinematic, graphic-poster, studio-product, and isometric-3d.
LUMIHOO_MODEL_PROFILES	unset	Optional JSON object defining model profiles. When set, it replaces the legacy single-model SGLANG_IMAGE_MODEL/IDEOGRAM_* profile behavior.
LUMIHOO_MODEL_PROFILE	unset	Optional active profile id override. Defaults to activeProfile in LUMIHOO_MODEL_PROFILES, then the first configured profile.

Model profiles

Use LUMIHOO_MODEL_PROFILES when different upstream models need different request shapes. Profiles are server-side deployment config; users do not choose models in the UI. The active profile controls the model, endpoint, prompt format, timeout, supported sizes, sampler preset menu, style preset menu, default seed, and extra upstream request fields.

{
  "activeProfile": "ideogram-v4",
  "profiles": [
    {
      "id": "ideogram-v4",
      "label": "Ideogram 4",
      "baseUrl": "http://localhost:30010/v1",
      "apiKeyEnv": "SGLANG_API_KEY",
      "model": "ideogram-ai/ideogram-4-nf4",
      "promptFormat": "ideogram-json",
      "timeoutMs": 110000,
      "sizes": ["1024x1024", "2048x2048"],
      "defaultSize": "1024x1024",
      "presets": [
        { "value": "V4_TURBO_12", "label": "Turbo" },
        { "value": "V4_DEFAULT_20", "label": "Balanced" },
        { "value": "V4_QUALITY_48", "label": "Quality" }
      ],
      "defaultPreset": "V4_DEFAULT_20",
      "stylePresets": [
        {
          "value": "natural",
          "label": "Natural",
          "prompt": ""
        },
        {
          "value": "brand-poster",
          "label": "Brand Poster",
          "prompt": "Bold graphic poster treatment with clear hierarchy, readable typography, limited spot colors, subtle print grain, and balanced negative space.",
          "styleDescription": {
            "aesthetics": "bold, graphic, high contrast, readable",
            "lighting": "flat even print lighting",
            "medium": "graphic_design",
            "art_style": "screenprint poster, bold display type, limited spot-color palette, subtle paper grain",
            "color_palette": ["#101114", "#F4EEE0", "#E34F35", "#1D8A99", "#F2B84B"]
          }
        }
      ],
      "defaultStylePreset": "natural",
      "extraBody": {}
    },
    {
      "id": "plain-image-model",
      "label": "Plain image model",
      "baseUrl": "http://localhost:30011/v1",
      "apiKeyEnv": "SGLANG_API_KEY",
      "model": "example/image-model",
      "promptFormat": "text",
      "sizes": ["1024x1024"],
      "defaultSize": "1024x1024",
      "presets": [],
      "stylePresets": [],
      "extraBody": { "quality": "high" }
    }
  ]
}

Supported promptFormat values are text, ideogram-json, and auto. Profiles with an empty presets array hide the preset control and omit preset from upstream requests. Profiles with an empty stylePresets array hide the style control. For ideogram-json profiles, selected styles are merged into style_description; for text profiles, selected styles are appended to the text prompt as a compact style instruction. extraBody cannot define reserved request keys: model, prompt, n, size, response_format, preset, or seed.

Curated profile examples live in model-profiles/.

Ideogram 4 settings

The default Ideogram 4 profile is tuned for interactive speed without dropping to draft quality:

Use case	Preset	Size	Notes
Fast drafts	V4_TURBO_12	1024x1024	Lowest latency. Use for exploration when detail is less important.
Normal app generation	V4_DEFAULT_20	1024x1024	Recommended default. About 40% of the quality preset steps with the same guidance shape.
Final-quality assets	V4_QUALITY_48	2048x2048	Highest fidelity. Use when waiting longer is acceptable.

For Ideogram 4 JSON prompts, Lumihoo keeps prompt content in the documented V4 caption contract and sends output dimensions through the upstream size field. Style presets are merged into style_description; user-supplied JSON prompt fields win over preset fields.

Batching is per prompt: the app sends one upstream /images/generations request with n equal to the selected count, then stores every safe returned image. Keep this path for same-prompt variants because it lets the backend batch the work and gives one timeout, one error response, and one storage transaction. Split into separate requests only when prompts, sizes, presets, or seeds differ.

Krea 2 Turbo

Krea 2 Turbo is the recommended Krea 2 inference profile. It uses the SGLang OpenAI-compatible images endpoint with model krea/Krea-2-Turbo, plain text prompts, 8 inference steps, no preset field, and square 1k-2k output sizes.

Start SGLang:

SGLANG_CACHE_DIT_ENABLED=true sglang serve \
  --model-path krea/Krea-2-Turbo \
  --num-gpus 1 \
  --port 30000

Load the profile for local development:

export LUMIHOO_MODEL_PROFILES="$(tr -d '\n' < model-profiles/krea-2-turbo.json)"
export SGLANG_API_KEY=EMPTY
pnpm dev

For Docker Compose, edit the profile baseUrl to http://host.docker.internal:30000/v1 before setting LUMIHOO_MODEL_PROFILES, or run SGLang as a Compose service on the same Docker network.

Development

Install dependencies:

pnpm install

Start the development server:

pnpm dev

Open http://localhost:3000.

Run linting:

pnpm lint

Run the full local check:

pnpm check

Create a production build:

pnpm build

Docker

Start the full stack:

docker compose up --build

Caddy serves the app at http://localhost:8080 and proxies image objects from MinIO at http://localhost:8080/images/....

Build the app image directly:

docker build -t ghcr.io/openhoo/lumihoo:local .

Released images are published to GitHub Container Registry:

• ghcr.io/openhoo/lumihoo:<version>
• ghcr.io/openhoo/lumihoo:<major>.<minor>
• ghcr.io/openhoo/lumihoo:sha-<commit>
• ghcr.io/openhoo/lumihoo:latest

Releases

Commits to main are linted as Conventional Commits by Hooversion. After CI passes, the release workflow creates the release commit, tag, and GitHub release, then builds and pushes the linux/amd64 Docker image to GHCR.

API

POST /api/generate accepts JSON:

{
  "prompt": "a great horned owl made of constellations",
  "count": 1,
  "preset": "V4_DEFAULT_20",
  "stylePreset": "cinematic",
  "size": "1024x1024",
  "seed": 12345
}

The route validates prompt length, image count, active-profile preset, active-profile style preset, active-profile size, and seed before forwarding the request to the configured SGLang endpoint. For ideogram-json profiles, plain text prompts are converted to structured JSON captions before they are sent upstream; already-JSON prompts are passed through as minified JSON with the selected style merged into style_description. Successful responses upload image bytes to MinIO, store metadata in Postgres, and return stable image URLs:

{
  "images": ["/images/generated/2026/07/07/936e30a8-c347-4d06-84dd-5ff8ed3542ac.png"],
  "items": [
    {
      "id": "936e30a8-c347-4d06-84dd-5ff8ed3542ac",
      "prompt": "a great horned owl made of constellations",
      "src": "/images/generated/2026/07/07/936e30a8-c347-4d06-84dd-5ff8ed3542ac.png",
      "createdAt": "2026-07-07T12:00:00.000Z"
    }
  ]
}

Share pages use GET /image?id=<image-id> and load image metadata from Postgres.

License

Apache License 2.0. See LICENSE.