CubicJ Cafe

AI Image-to-Video generation web frontend powered by ComfyUI.

Upload an image, write a prompt, and get a generated video delivered to Discord — with a queue, multi-model routing, and a full admin dashboard on top.

Screenshots

Generation	Generation form
Settings	Admin dashboard

Features

• Image-to-Video — Upload an image with a prompt, get a generated video

• Multi-Model — Six workflow pipelines dispatched through a single capability-driven registry

  Model	End Image	Audio Output	Duration
  WAN 2.2	✓	—	5–7s
  LTX (Anime)	—	✓	5–7s
  LTX (Real)	✓	✓	5–7s
  L+W (LTX + WAN hybrid)	✓	✓	5–8s
  H3 FL2VA (MiniMax)	✓	✓ (native)	frame-count based
  H3 Ref2VA (MiniMax)	—	✓ (native)	frame-count based

  H3 FL2VA takes a first frame, a last frame, or both — any single image is enough (F2VA / L2VA / FL2VA modes).

  H3 Ref2VA accepts up to 9 reference images, 3 reference videos, and 3 reference audio clips. Output aspect ratio comes from the first reference image or a custom ratio.

• Queue System — Serializable queue with atomic position assignment and real-time status tracking

• Reference Inputs — Drag-and-drop image, video, and audio uploads with per-kind limits (9 / 3 / 3) and first-image or custom aspect-ratio control

• LoRA Presets & Bundles — Drag-and-drop preset management, per-model availability gated by capability flags

• Audio Presets — Upload and reuse audio clips as generation input on the LTX-family models; H3 Ref2VA takes reference audio directly

• Prompt Translation — Built-in translation endpoint for non-English prompts

• Discord Integration — OAuth2 auth + in-process bot delivery of completed videos

• Admin Dashboard — Model activation toggles, queue control and pause scheduling, live log viewer (SSE), ComfyUI monitoring, DB browser

• Multi-Server — Auto-selects between local and cloud (RunPod) ComfyUI instances

Tech Stack

Layer	Technology
Frontend	Next.js 16, React 19, TailwindCSS 4, Shadcn/ui; IBM Plex Sans KR + self-hosted IBM Plex Mono; light-only warm token theme (DESIGN.md)
Backend	Next.js API Routes, Prisma 7 (better-sqlite3 driver adapter), SQLite
Auth	Custom Discord OAuth2 (HttpOnly cookie sessions)
Validation	Zod v4 schemas across all route handlers
AI Backend	ComfyUI API (JSON workflow graphs)
Bot	Discord.js (in-process)
Testing	Vitest, SQLite test DB, route handler direct invocation
Logging	Custom unified logger (server + client → files + SSE)
Deployment	systemd + Nginx + SSL (Next.js standalone)

Project Structure

src/
├── app/
│   ├── api/          # 46 API route handlers
│   │   ├── i2v/      # Video generation endpoint
│   │   ├── queue/    # Queue management + monitoring
│   │   ├── admin/    # Protected admin routes
│   │   └── auth/     # Discord OAuth2 flow
│   ├── i2v/          # Generation page
│   ├── admin/        # Admin dashboard
│   ├── profile/      # User profile
│   └── settings/     # User settings
├── lib/
│   ├── comfyui/      # Queue/job monitors, server management
│   │   └── workflows/  # Per-model builders (wan, ltxa, ltxr, ltx-wan, h3-fl2va, h3-ref2va) + registry
│   ├── database/     # Prisma service layer
│   ├── auth/         # Session management, withAuth/withAdmin HOF
│   ├── validations/  # Zod schemas + parse helpers
│   └── logger.ts     # Client-safe unified logger
├── components/       # Shadcn/ui base + domain components
├── contexts/         # React Context (session, form state)
└── hooks/            # Custom React hooks
tests/                # Mirrors src/ (route handler tests hit real DB sessions)
├── app/api/          # API route tests via direct handler invocation
├── lib/              # Module tests
├── helpers/          # Shared fixtures, seeds, auth builders
└── perf/             # Benchmarks

Development

npm install
npm run prisma:migrate    # Set up database
npm run dev               # Dev server
npm test                  # 666 tests
npm run type-check        # tsc --noEmit
npm run lint              # ESLint

Deployment

Runs as a systemd service with Next.js standalone output. Environment variables live in .env (not version controlled — see .env.example for the template).

Production uses .env.production loaded via systemd EnvironmentFile.

npm run build
# Deploy .next/standalone + .next/static + public to target
# systemd EnvironmentFile loads .env.production
# ExecStart: node .next/standalone/server.js

Architecture Notes

• ComfyUI workflows are JSON node graphs — each model has a completely different structure, dispatched via workflow-router.ts
• Model capabilities (end image, audio, duration, LoRA) live in a single registry that gates both UI and API validation
• Queue uses Serializable isolation for atomic position assignment
• Logger is split: logger.ts stays client-safe (no fs), file I/O in logger-file.ts connected via instrumentation.ts
• Discord bot runs in-process (not a separate service)
• All state singletons use globalThis to survive Next.js dev hot reload

License

MIT