Category
Local (ComfyUI)
LTX23.net
Text-to-Video (Pro)
Synced audio + LTXVScheduler for tight pacing
- Best for cinematic prompts and structured camera direction
- Includes synced audio pass (when enabled)
- Optimized defaults for 1080p preview runs
LTX 2.3 ComfyUI Mastery
Native 9:16 portrait support, sharper VAE textures, and Gemma 3 integration. The definitive guide to professional local video generation.
Instant Access
No GPU? No Problem.
Skip the Python environment and VRAM errors. Run LTX 2.3 in the cloud with our optimized high-speed generator.
Before you install
Technical prerequisites
Get the baseline right first. Most failures (black frames, OOM) come from mismatched versions or missing node packs.
ComfyUI
v0.16.1 or higher (update via Manager).
GPU VRAM
12GB+ (GGUF) or 32GB+ for the 22B dev checkpoint.
Custom nodes
Install `ComfyUI-LTXVideo` (search “LTXVideo” in Manager).
Setup
Install LTX 2.3 in ComfyUI (step-by-step)
Follow this exact order to avoid missing nodes or incorrect model discovery.
- 01
Update ComfyUI via Manager
Open ComfyUI-Manager → Update → restart ComfyUI after it finishes.
- 02
Install the LTXVideo custom nodes
Manager → Search → install “LTXVideo” (the `ComfyUI-LTXVideo` node pack).
- 03
Set model paths
Checkpoints: `ComfyUI/models/checkpoints/`. Text encoders: place into the folder expected by your LTXVideo nodes (then restart).
Workflows
Optimized workflow downloads
Start local with these optimized JSON workflows, or run the same ideas online in one click.
T2V
Image-to-Video (Cinematic)
Updated VAE for stable texture + reduced flicker
- Ideal for product shots, storyboard frames, and concept art
- Keeps composition while adding motion and parallax
- Safer motion strength defaults to avoid jitter
Low‑VRAM GGUF
12GB/16GB-friendly config using quantized weights
- Designed for 12GB+ cards (quality vs speed trade-off)
- Recommended for local iteration and prompt testing
- Lower resolution first, then upscale in stages
Hardware reality check
22B local runs can be expensive (VRAM warning)
If your GPU can’t comfortably run the big checkpoints, don’t churn on OOM errors—use the cloud generator and keep creating.
VRAM guidance
- 12GB+ VRAM: consider GGUF/quantized weights for iteration.
- 32GB+ VRAM: more comfortable for heavier checkpoints and higher-res runs.
Convenience factor
Local vs. LTX23.net (what you gain)
Local is powerful when you have the hardware. The web generator is built for speed, stability, and sharing.
Setup & maintenance
Python env, node packs, model paths, frequent updates.
1-click, no local setup.
Performance
Hardware dependent; can be slow on mid-range GPUs.
Instant GPU access + consistent rendering.
Quality upgrades
Manual upscaling + pipeline tuning.
4K upscaling and optimized defaults.
Sharing
Local files and manual export flows.
Shared showcase and easy sharing.
Prompting
Director-style prompting guide
LTX 2.3 responds best when you describe shot intent like a director: framing, subject action, environment, then light.
The formula
[Shot Type] + [Subject] + [Action] + [Environment] + [Lighting]
Sample LTX 2.3 prompt
Wide shot, a lone motorcyclist rides through a neon-lit rainy street at night, reflections on wet asphalt, slow tracking shot, cinematic lighting, subtle film grain, high detail textures.
Upgrade
LTX 2.0 vs LTX 2.3 (what actually changes)
If you're upgrading an older workflow, these are the three differences you’ll feel immediately.
Category
LTX 2.0
LTX 2.3
Resolution & aspect
Mostly landscape-first setups; vertical needs extra handling.
Native 9:16 portrait support, easier mobile-first output.
VAE quality
More texture softness; higher chance of shimmer.
Sharper textures, cleaner edges, reduced flicker in motion.
Text adherence
Good, but can drift with complex camera/lighting cues.
Stronger adherence (Gemma 3 integration) with director-style prompts.
Troubleshooting
FAQ: black frames, OOM, missing nodes
The three most common local issues and the fastest fixes.
Black frames
Usually a model path mismatch or an incompatible VAE/text encoder. Recheck checkpoint placement, restart ComfyUI, and verify the text encoder is discovered by the nodes.
Out of Memory (OOM)
Lower resolution first, reduce batch/frames, switch to GGUF/quantized weights, and close other GPU apps. If you’re on 12–16GB VRAM, avoid 22B dev runs.
Missing nodes
Open Manager → ensure `ComfyUI-LTXVideo` is installed and updated. If nodes still don’t appear, restart and check for extension load errors in the console.
Want more examples and prompt patterns?