---

name: ltx-video description: This skill should be used when the user asks to "generate video with LTX-2", "create video from text", "animate an image", "use GAMMA for video generation", or needs local AI video generation on the GAMMA GPU server. Covers text-to-video, image-to-video, and auto-cataloging workflows with the LTX-2 19B model. version: 1.0.0 license: MIT

LTX-2 Video Generation

Generate AI videos using the LTX-2 19B model running on the GAMMA GPU server.

Quick Reference

---

Architecture

CopyGAMMA (Linux GPU Server)          ALPHA (Current Machine)
┌────────────────────┐            ┌────────────────────┐
│  LTX-2 API :8001   │◄──────────►│  ltx2_client.py    │
│  • RTX 4090        │  Tailscale │  • Job submission   │
│  • 24GB VRAM       │            │  • Status polling   │
│  • FP8 inference   │            │  • Auto-cataloging │
└────────────────────┘            └────────────────────┘

---

Generation Workflow

1. Text-to-Video

Copyfrom ltx2_client import LTX2Client

client = LTX2Client()

# Check server health
health = client.health()
print(f"GPU: {health.gpu}, Active Jobs: {health.active_jobs}")

# Submit generation job
job = client.text_to_video(
    prompt="A serene mountain lake at sunrise, mist rising from the water, cinematic drone shot",
    negative_prompt="blurry, shaky, overexposed",
    num_frames=121,  # ~4.8 seconds at 25fps
    seed=12345,
)
print(f"Job submitted: {job.job_id}")

# Wait for completion and download
path = client.download(job.job_id)
print(f"Video saved: {path}")

2. Image-to-Video (Animation)

Copy# Animate a static image
job = client.image_to_video(
    image_path="/path/to/image.jpg",
    prompt="Camera slowly zooms out revealing the full scene",
    num_frames=121,
    seed=42,
)

3. Complete Workflow with Progress

Copydef on_progress(stage, job):
    print(f"[{stage}] {job.job_id}: {job.status}")

path = client.generate_and_download(
    prompt="Futuristic cityscape with flying vehicles",
    progress_callback=on_progress,
    num_frames=121,
    seed=42,
)

---

CLI Usage

Copy# Check server health
python ltx2_client.py health

# Generate video (submit only)
python ltx2_client.py generate "A beautiful sunset over the ocean" --frames 121

# Generate and wait for download
python ltx2_client.py generate "Mountain landscape with clouds" --wait --output my_video.mp4

# Check job status
python ltx2_client.py status <job_id>

# Download completed job
python ltx2_client.py download <job_id> --output video.mp4

# List recent jobs
python ltx2_client.py list --limit 10

---

Generation Parameters

Duration Reference

---

Prompt Writing Guide

Effective Prompts

CopyGood: "A golden retriever running through a sunlit meadow, slow motion,
      shallow depth of field, cinematic 4K"

Bad:  "dog running"

Structure

1. Subject: What is the main focus
2. Action: What is happening
3. Setting: Where it takes place
4. Style: Cinematic, documentary, animated, etc.
5. Technical: Camera movement, lighting, quality modifiers

Style Keywords

Negative Prompts

Common negative prompts to improve quality:

Copyblurry, low quality, distorted, watermark, text, logo,
overexposed, underexposed, shaky, jittery, artifacts,
morphing faces, unnatural motion

---

Auto-Cataloging

Generated videos are automatically cataloged to MediaDB:

Copyfrom ltx2_client import LTX2Client
from media_db import MediaDB
from embedding_service import EmbeddingService

# Generate video
client = LTX2Client()
path = client.generate_and_download(prompt="...")

# Catalog with metadata
db = MediaDB()
embedding = EmbeddingService().embed(prompt)

record = {
    "id": generate_id(),
    "filename": path.name,
    "filepath": str(path),
    "media_type": "video",
    "description": prompt,
    "quality_class": 1,  # AI-generated = high quality
    "style_tags": ["ai_generated", "ltx2"],
    "vector": embedding,
    # ... additional metadata
}
db.add_record(record)

---

API Reference

Health Check

Copycurl https://gamma.tail5f2bae.ts.net:8001/health

Response:

Copy{
  "status": "healthy",
  "model": "ltx-2-19b-dev-fp8",
  "pipeline_loaded": true,
  "cuda_available": true,
  "gpu": "NVIDIA RTX 4090",
  "active_jobs": 0
}

Text-to-Video

Copycurl -X POST https://gamma.tail5f2bae.ts.net:8001/api/v1/text-to-video \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "A futuristic cityscape with flying vehicles",
    "num_frames": 121,
    "seed": 42
  }'

Job Status

Copycurl https://gamma.tail5f2bae.ts.net:8001/api/v1/status/{job_id}

Download

Copycurl -o video.mp4 https://gamma.tail5f2bae.ts.net:8001/api/v1/download/{job_id}

---

Troubleshooting

Server Not Responding

Copy# Check if GAMMA is reachable
tailscale ping gamma

# Check API directly
curl -k https://gamma.tail5f2bae.ts.net:8001/health

GPU Out of Memory

Reduce parameters:

• Lower num_frames (try 61 instead of 121)
• Use smaller resolution (512x384 instead of 768x512)

Generation Quality Issues

1. Use more specific prompts with style keywords
2. Add comprehensive negative prompts
3. Increase steps (40 → 50) for better quality
4. Adjust cfg_scale (higher = more prompt adherence)

---

Integration with Other Skills

With Gemini Super Gems (Veo)

Use LTX-2 for quick iterations, Veo for final production:

Copy# Quick prototype with LTX-2
ltx_path = ltx_client.generate_and_download(prompt)

# Review and refine...

# Final production with Veo (higher quality)
# Use Super Gems browser automation

With LM Studio (Embeddings)

Generate embeddings for semantic search:

Copyfrom embedding_service import EmbeddingService

embedding = EmbeddingService().embed(prompt)
# Use for semantic search in MediaDB

---

Bundled Resources

• scripts/generate.py - Text-to-video generation
• scripts/animate.py - Image-to-video animation
• scripts/catalog.py - Auto-cataloging with embeddings
• references/parameters.md - Full parameter documentation
• references/prompt-guide.md - Prompt writing best practices