Reigh-Worker

GPU worker for Reigh — handles video generation tasks using Wan2GP.

This is a headless wrapper around the Wan2GP engine that enables programmatic video generation and queue-based processing. It connects to the Reigh application to claim and execute video generation tasks on GPU infrastructure.

Architecture

• Reigh — The main application (frontend + API)
• Reigh-Worker (this repo) — GPU worker that processes video generation jobs using Wan2GP

Powered by Wan2GP

This worker is built on top of Wan2GP by deepbeepmeep, a powerful video generation engine. The Wan2GP/ directory contains the upstream engine code with some modificiations.

---

Running With Reigh: Quick Start (GPU)

1. Create venv: python3 -m venv venv_headless_wan2gp && source venv_headless_wan2gp/bin/activate
2. Install deps:
   • python -m pip install --upgrade pip wheel setuptools
   • python -m pip install -r Wan2GP/requirements.txt
   • python -m pip install -r requirements.txt
3. Set credentials and run the worker:

SUPABASE_SERVICE_ROLE_KEY="your-service-role-key-from-reigh" python worker.py \
    --supabase-url "https://your-project.supabase.co" \
    --worker "my-worker-001"

Note: Get your credentials from reigh.art. For standalone usage without credentials, see Running Without Reigh below.

Running Without Reigh: Standalone Usage

You can use the generation engine directly without connecting to Reigh's database. This is useful for local testing, scripting, or building custom pipelines.

Using the Examples

The examples/ directory contains self-contained scripts:

# Join two video clips with a smooth AI-generated transition
python examples/join_clips_example.py \
    --clip1 scene1.mp4 \
    --clip2 scene2.mp4 \
    --output transition.mp4 \
    --prompt "smooth camera glide between scenes"

# Regenerate a range of frames within a video (fix corrupted frames, etc.)
python examples/inpaint_frames_example.py \
    --video my_video.mp4 \
    --start-frame 45 \
    --end-frame 61 \
    --output fixed_video.mp4 \
    --prompt "smooth continuous motion"

Using HeadlessTaskQueue Directly

For custom scripts, use HeadlessTaskQueue to manage model loading and generation:

import os, sys
from pathlib import Path

# Setup paths
project_root = Path(__file__).parent
sys.path.insert(0, str(project_root))
wan_root = project_root / "Wan2GP"

from headless_model_management import HeadlessTaskQueue, GenerationTask

# Initialize queue (loads models on first use)
queue = HeadlessTaskQueue(wan_dir=str(wan_root), max_workers=1)
queue.start()

try:
    # Create a generation task
    task = GenerationTask(
        id="my_task_001",
        model="wan_2_2_vace_lightning_baseline_2_2_2",
        prompt="a cat walking through a garden",
        parameters={
            "video_length": 81,
            "resolution": "896x512",
            "num_inference_steps": 6,
            "guidance_scale": 3.0,
            "seed": 42,
        }
    )
    
    # Submit and wait for result
    queue.submit_task(task)
    result = queue.wait_for_completion(task.id, timeout=600)
    
    if result.get("success"):
        print(f"Generated: {result['output_path']}")
    else:
        print(f"Failed: {result.get('error')}")
finally:
    queue.stop()

Available Models

Common model presets (see Wan2GP/defaults/*.json for full list):

Model Key	Type	Description
wan_2_2_vace_lightning_baseline_2_2_2	VACE	Fast video inpainting/editing (6 steps)
wan_2_2_i2v_lightning_baseline_2_2_2	I2V	Image-to-video with Lightning LoRAs
wan_2_2_t2v_14B	T2V	Text-to-video (14B params)
qwen_image_edit_20B	Image	Qwen-based image editing

Helper Scripts

• scripts/gpu_diag.sh – Prints GPU/NVML diagnostics to help debug nvidia-smi issues in containers
• debug.py – Debug tool for investigating tasks and worker state
• create_test_task.py – Create test tasks in Supabase for debugging

Debugging

Use python debug.py to investigate tasks and worker state:

python debug.py task <task_id>          # Investigate a specific task
python debug.py tasks --status Failed   # List recent failures

Notes

• The wrapper prefers upstream‑first fixes to maintain clean Wan2GP updates.