LTX-2.3 Video Generation

Generate ~5 second video clips from text prompts or images using the LTX-2.3 22B DiT model. Runs on Modal (A100-80GB). Requires MODALLTX2ENDPOINT_URL in .env.

Quick Reference

# Text-to-video
uv run tools/ltx2.py --prompt "A sunset over the ocean, golden light on waves, cinematic" --output sunset.mp4

# Image-to-video (animate a still image)
uv run tools/ltx2.py --prompt "Gentle camera drift, soft ambient motion" --input photo.jpg --output animated.mp4

# Custom resolution and duration
uv run tools/ltx2.py --prompt "..." --width 1024 --height 576 --num-frames 161 --output wide.mp4

# Fast mode (fewer steps, quicker)
uv run tools/ltx2.py --prompt "..." --quality fast --output quick.mp4

# Reproducible output
uv run tools/ltx2.py --prompt "..." --seed 42 --output reproducible.mp4

Parameters

Style LoRAs

Style LoRAs bias the output toward a specific visual aesthetic. They're baked into the Modal image and selected per-request; switching LoRAs forces a pipeline rebuild (~60s one-time cost per container lifetime per switch).

crt-terminal — CRT / pixel-art terminals

Base: LTX-2.3 22B, trained by @lovis93 (Apache 2.0).

# Trigger word is auto-prepended — write the prompt normally
uv run tools/ltx2.py --lora crt-terminal \
  --prompt "a terminal typing out \"\\$ claude --continue\" character by character in glowing green pixel font, scanlines, phosphor glow, low choppy frame rate, hacker mood" \
  --output crt_claude.mp4

What the preset changes:

• Prepends crtanim, to the prompt (the LoRA's trigger word)
• Defaults to 1024×1024, 121 frames (the ratio it was trained on)
• Relaxes the default negative prompt so on-screen text isn't filtered out

Prompt pattern: <CRT aesthetic> → <color palette> → <animation style> → <subject> → <literal text in quotes> → <mood>. Keep on-screen text to 1–3 words — the model can't render long strings reliably. The LoRA prefers static framing; ask for camera moves explicitly if you want them.

Valid Frame Counts

(n - 1) % 8 == 0: 25 (~1s), 49 (~2s), 73 (~3s), 97 (~4s), 121 (~5s default), 161 (~6.7s), 193 (~8s max practical).

Common Resolutions

Prompting Guide

LTX-2 responds well to cinematographic descriptions. Layer these dimensions:

• Camera: "Slow dolly forward", "Aerial drone shot", "Tracking shot", "Static wide angle"
• Lighting: "Golden hour", "Cinematic lighting", "Neon-lit", "Soft diffused light"
• Motion: "Timelapse of...", "Slow motion", "Gentle camera drift", "Gradually transitions"
• Style: "Shot on 35mm film", "Documentary style", "Clean minimal aesthetic"
• Negative: Always implicitly avoids "worst quality, blurry, jittery, watermark, text, logo"

Keep prompts under 200 words. Be specific about the scene.

Good Prompts

# Atmospheric b-roll
"Aerial drone shot slowly flying over turquoise ocean waves breaking on white sand, golden hour sunlight, cinematic"

# Product/tech scene
"Close-up of hands typing on a mechanical keyboard, shallow depth of field, soft desk lamp lighting, cozy atmosphere"

# Abstract background
"Dark moody abstract background with flowing blue light streaks, subtle geometric grid, bokeh particles floating, cinematic tech atmosphere"

# Animate a portrait
"Professional headshot, subtle natural head movement, confident warm expression, studio lighting, shallow depth of field"

# Animate a slide/screenshot
"Gentle subtle particle effects floating across a presentation slide, soft ambient light shifts, very slight camera drift"

Bad Prompts

# Too vague
"A cool video"

# Too many competing ideas
"A cat riding a skateboard while juggling fire on the moon during a thunderstorm"

# Describing text/UI (model can't render text reliably)
"A website showing the text 'Welcome to our platform'"

Video Production Use Cases

B-Roll Clips

Generate atmospheric 5s shots for cutaways between narrated scenes:

uv run tools/ltx2.py --prompt "Futuristic holographic interface, glowing data visualizations, clean workspace, cinematic" --output broll_tech.mp4
uv run tools/ltx2.py --prompt "Aerial view of European city at golden hour, modern architecture" --output broll_europe.mp4

Animated Slide Backgrounds

Feed a slide screenshot and add subtle motion:

uv run tools/ltx2.py --prompt "Gentle particle effects, soft ambient light shifts, very slight camera drift" --input slide.png --output animated_slide.mp4

Animated Portraits

Bring still headshots to life:

uv run tools/ltx2.py --prompt "Subtle natural head movement, warm expression, professional lighting" --input headshot.png --output animated_portrait.mp4

Stylized Character Cameo (SadTalker Alternative)

For non-realistic faces — fantasy characters, masked figures, heavy beards, helmets, illustrations — SadTalker often produces uncanny or broken lip sync because it's trained on photoreal humans. LTX-2 image-to-video is frequently a better choice when lip-sync precision isn't critical (the viewer's brain fills in the gap as long as something is moving). Prompt for motion + atmosphere, not phonemes:

uv run tools/ltx2.py \
  --input character_portrait.png \
  --prompt "Ancient warrior speaks slowly with gravitas, beard shifts subtly, glowing aura pulses, embers drift past, slow head movement, cinematic close-up, mystical atmosphere" \
  --width 768 --height 768 \
  --output character_speaking.mp4

When LTX-2 wins over SadTalker:

• Stylized / illustrated / fantasy characters
• Heavy facial hair or accessories obscuring the mouth
• Masked or helmeted figures
• Short cameo lines where atmosphere matters more than precision
• Dramatic VO rather than dialogue

When SadTalker still wins:

• Photoreal human presenters
• Full sentences where mouth shape needs to match phonemes
• Tutorials / talking-head explainers where the viewer is effectively reading lips

Branded Intro/Outro

Generate abstract motion backgrounds for title cards:

uv run tools/ltx2.py --prompt "Dark moody background with flowing blue and coral light streaks, bokeh particles, cinematic tech atmosphere, no text" --output intro_bg.mp4

Combining with Other Tools

LTX-2 generates raw clips. Combine with the rest of the toolkit:

Technical Details

• Model: LTX-2.3 22B DiT (Lightricks), bf16
• GPU: A100-80GB on Modal (~$4.68/hr)
• Inference: ~2.5 min per clip (768x512, 121 frames, 30 steps)
• Cost: ~$0.20-0.25 per 5s clip
• Cold start: ~60-90s (loading ~55GB weights)
• Output: H.264 MP4 with synchronized ambient audio (24fps)
• Max duration: ~8s (193 frames) per clip

Known Limitations

• Training data artifacts: ~30% of generations may have unwanted logos/text from training data. Re-run with different --seed.
• Text rendering: Cannot reliably generate readable text in video. Use Remotion overlays instead.
• Max duration: ~8s per clip. Longer content needs stitching.
• Audio: Generated audio is ambient/environmental only. Use voiceover/music tools for speech and music.
• License: Community License — free under $10M revenue, commercial license needed above that.

Setup

# 1. Create Modal secret for HuggingFace (one-time)
uv run modal secret create huggingface-token HF_TOKEN=hf_your_token

# 2. Deploy (downloads ~55GB of weights, takes ~10 min)
uv run modal deploy docker/modal-ltx2/app.py

# 3. Save endpoint URL to .env
echo "MODAL_LTX2_ENDPOINT_URL=https://yourname--video-toolkit-ltx2-ltx2-generate.modal.run" >> .env

# 4. Test
uv run tools/ltx2.py --prompt "A candle flickering on a dark table, cinematic" --output test.mp4

Important: HuggingFace token needs read-access scope. Accept the Gemma 3 license before deploying. Unauthenticated downloads are severely rate-limited.