🪄 Open Fantasia — Local Image & Video Generation

Generate images locally on your GPU using FLUX.1-schnell or Stable Diffusion, and short videos with Wan2.1 — no external API, no cost per image or clip.

---

Gallery

Generated locally with FLUX.1-schnell, low quality preset. Speed varies by GPU.

---

Video Gallery

Short clips generated locally with Wan2.1 (Wan-AI/Wan2.1-T2V-1.3B, mid quality):

Generated locally with Wan2.1, mid quality preset. ~5s clips at 480P.

---

Features

• 🚀 Local generation — no API keys, no cloud, no per-image cost
• ⚡ FLUX.1-schnell — fast distilled model, great quality
• 🌸 FLUX.2-klein — lighter experimental model
• 🎨 Stable Diffusion 1.5 — classic fallback, runs on 6GB VRAM
• ⚡ SD-Turbo — near-instant generation (~1-3s at 512×512)
• 🎬 Wan2.1 text-to-video — short MP4 clips from text (~8GB VRAM)
• 🔢 Batch generation — generate 1–4 images per prompt (different seeds)
• 💬 OpenClaw slash command — /fantasia <prompt> with inline buttons
• 🔁 Persistent server — model stays in VRAM, near-instant after first load

---

Video Generation (Wan2.1)

Generate a short MP4 clip from a text prompt using Wan-AI/Wan2.1-T2V-1.3B (~8.19 GB VRAM, 480P, ~4 min for a 5s clip on an RTX 4090):

curl -X POST http://localhost:8765/video \
  -H "Content-Type: application/json" \
  -d '{"prompt":"a cat walking through a neon city at night","quality":"mid"}'

Returns raw video/mp4 bytes. Saved to ~/.openclaw/media/fantasia/videos/<timestamp>.mp4.

Model aliases: wan, wan13, wan2.1, wan-1.3b → Wan-AI/Wan2.1-T2V-1.3B

VRAM note: The server auto-unloads the image/Qwen pipelines before loading Wan2.1 to free VRAM, and reloads them on the next image request.

---

Requirements

• Python 3.10+
• CUDA GPU — 8GB VRAM minimum (16GB recommended for FLUX)
• HuggingFace account + HF_TOKEN in .env (see .env.example)
• ~16GB disk space for FLUX.1-schnell (downloaded once)

---

Quick Start

git clone https://github.com/fabiopacifici-bot/open-fantasia-imagegen
cd open-fantasia-imagegen
cp .env.example .env   # add your HF_TOKEN
bash setup.sh

setup.sh will:

1. Create a Python venv and install dependencies
2. Download the recommended model (~16GB, one-time)
3. Install and enable the systemd service
4. Start the server at http://localhost:8765

---

Usage

Via OpenClaw slash command (recommended)

/fantasia a samurai cat at sunset
/fantasia cyberpunk city --quality high
/fantasia quick sketch --quality low --count 3
/fantasia portrait --model klein
/fantasia setup        ← first-time setup

Inline buttons: Call /fantasia with no prompt to get model/quality/count selection buttons.

Direct API

# Generate an image
curl -X POST http://localhost:8765/generate \
  -H "Content-Type: application/json" \
  -d '{"prompt": "a cat in space", "quality": "low", "enhance": true, "count": 1}'

# Health check
curl http://localhost:8765/health

CLI

.venv/bin/python src/imagegen.py \
  --prompt "a samurai cat at sunset" \
  --quality low \
  --output output.png

---

Quality Presets

Note: Speed varies significantly by GPU. FLUX runs in BF16 precision. Use torchao autoquant to reduce VRAM usage and improve throughput.

---

Models

---

Server Management

# Start / stop / restart
systemctl --user start fantasia.service
systemctl --user stop fantasia.service
systemctl --user restart fantasia.service

# Status & logs
systemctl --user status fantasia.service
journalctl --user -u fantasia.service -f

Output images are saved to ~/.openclaw/media/fantasia/.

---

Project Structure

open-fantasia-imagegen/
├── src/
│   ├── server.py       # FastAPI inference server
│   └── imagegen.py     # Pipeline loader + generation logic
├── assets/             # README demo images
├── setup.sh            # First-time setup script
├── .env.example        # Environment variable template
├── requirements.txt
└── .specs/             # Plans, docs, debugging notes

---

License

MIT

Roadmap

v1.3 — Turbo mode ✅

• sd-turbo available as --model turbo (stabilityai/sd-turbo)
• ~1-3s generation at 512×512 on CUDA

v2.0 — Image Editing

• Instruction-based image editing via Qwen2.5-VL
• New endpoint: POST /edit — { "image": "<path>", "instruction": "make the sky purple" }
• Slash command: /fantasia edit <image> <instruction>