LTX 2.3: AI Video Generation Locally on RTX 5090 — What Works, What Doesn't

Four days ago, Lightricks released LTX 2.3 — the latest version of their open-source video model. I pulled the model onto my RTX 5090 immediately and started testing.

The result: a showreel made entirely of LTX 2.3 generated shots. No post-processing, no cherry-picking through hundreds of attempts. Just prompt, render, done.

Fully AI-generated showreel with LTX 2.3 — running locally on RTX 5090

What is LTX 2.3?

LTX 2.3 is an open-source model from Lightricks (the company behind LTX Studio) with 22 billion parameters. It generates video and audio in a single pass — no stitching separate models together after the fact.

Key specs:

The big differentiator: LTX 2.3 is the only model in its class that runs fully locally. No API calls, no cloud, no wait times. Download the weights, start it up, prompt away.

Why Open Source Matters Here

Runway Gen-4.5, Sora, Kling — all solid models. But all closed source and API-only. That means:

• Per-clip costs add up fast
• No control over data and prompts
• Vendor lock-in

LTX 2.3 runs on my RTX 5090 at ~63 seconds render time per 6-second clip. That's roughly 57 clips per hour at zero API cost. A batch of 50 clips on Runway would easily run $50+. Here: electricity.

Available on HuggingFace with day-0 ComfyUI support and a standalone desktop app.

What LTX 2.3 Does Really Well

After dozens of prompts and variations, a clear picture emerges. LTX 2.3 has genuine strengths — if you know how to play to them.

Vehicles in Motion

Cars from behind or the side at speed — this lands almost every time. Motion blur is convincing, proportions stay stable.

Porsche on a rain-soaked highway — a sweet spot for LTX 2.3

Mood and Light

Everything involving golden hour, backlight, and silhouettes looks excellent. The model understands cinematic lighting language and produces consistent, filmic atmosphere.

Rooftop scene with skyline — backlight and silhouettes are a clear strength

Cinematic Camera Movement

Slow orbits, drone shots, and following people from behind work reliably. The model keeps the motion smooth and perspective stable.

Bamboo forest in Kyoto — following shot from behind

The Prompt Language That Works

What consistently delivers good results:

• 35mm Kodak film grain — the model speaks film stock language fluently
• Anamorphic lens + shallow DOF — consistent bokeh
• Real location names — "Kyoto bamboo forest" or "Dubai Marina" outperforms generic descriptions
• Wet surfaces, rain, reflections — light play on water often looks stunning
• Single subject, single action — simplicity wins

Model in the desert — single subject with shallow depth of field

Where LTX 2.3 Still Struggles

No model is perfect, and LTX 2.3 has clear limitations. Knowing them saves frustration:

Problem Areas

• Text, logos, license plates — turn into illegible mush. No current model handles this well, but it's especially noticeable here
• Faces at medium distance — tight close-ups of weathered faces work, but faces in medium shots distort
• Hands and fingers in close-up — the classic AI problem lives on
• Small props (cigars, pens, phones) — proportions and details are rarely right
• Running/swimming motion — physics break down, unnatural loops
• Crowds with visible faces — silhouettes work, details don't
• Food close-ups — textures and consistency feel uncanny
• Sepia/vintage color casts — the model drifts with heavy color grading
• "Stare into camera" — eye contact is inconsistent
• Harsh single-source light — shadows and highlights become unpredictable
• Cultural decorative details (flags, signs, ornaments) — get simplified or distorted
• Fade to black — the model likes to add this uninstructed

What's New in Version 2.3

What improved over previous versions:

• New VAE — sharper output, better texture and edge preservation
• Cleaner audio — less background noise and artifacts
• Prompt adherence — rebuilt text connector, the model follows complex prompts more closely
• Native portrait video — 9:16 without cropping from landscape
• Last-frame interpolation — provide a first and last frame, the model generates the video in between
• Image-to-video — more natural motion, fewer static clips

My Setup

For anyone who wants to try it:

• GPU: NVIDIA RTX 5090
• Model: ltx-2.3-22b-distilled (8-step variant for faster inference)
• Render time: ~63 seconds per 6-second clip
• Download: HuggingFace or GitHub
• Requirements: Python 3.12+, CUDA 12.7+, PyTorch 2.7

The distilled variant needs only 8 diffusion steps instead of the full sampling of the dev model. Quality difference is minimal, speed gain is massive.

Verdict: A Real Tool for Video Producers

LTX 2.3 is no longer a toy. It's a production tool with real strengths — when you work within its boundaries.

For mood shots, B-roll, establishing shots, and atmospheric scenes, it's production-ready now. Combined with a solid edit and real footage, this creates a workflow that was unthinkable just a year ago.

The fact that everything runs locally, is open source, and incurs zero API costs makes it particularly interesting for small studios and freelancers.

More testing to come — I'm curious what's possible with custom LoRAs and the upscaler.

Want to use AI-generated video for your business? I help with setup, workflow, and integrating it into your production pipeline. Let's talk →