LTX Desktop: Run LTX 2.3 Locally Without ComfyUI

If you have been avoiding LTX 2.3 because the local setup means a 46 GB checkpoint, a separate 12B text encoder, and a ComfyUI node graph — there is an official option most people never find.

LTX Desktop is Lightricks' open-source desktop app for generating video with LTX models locally. It runs on Windows and Linux with an NVIDIA GPU, and — this is the part that surprises people — on Apple Silicon Macs via MPS. When your hardware falls below the thresholds, it transparently switches to an API mode instead of failing.

This guide covers who can run it locally, what the free API key actually covers, how its LoRA handling works, and one genuinely counterintuitive gotcha in how it decides your Mac is "too small." All from the LTX-Desktop repository, checked July 2026.

Status note straight from the repo: LTX Desktop is Beta, and the README warns to expect breaking changes. Treat it as a fast-moving app, not a frozen product.

Can your machine run it locally?

The repository publishes an explicit hardware table:

Note that 16 GB VRAM number. It is the closest thing Lightricks publishes to a consumer hardware floor for LTX 2.3 — the Python package documents no inference minimum at all. If you have been trying to work out whether your card is enough, this table is the most authoritative signal available.

Full system requirements for local generation:

• Windows: Windows 10/11 x64, NVIDIA CUDA GPU ≥16 GB VRAM, 16 GB+ system RAM (32 GB recommended), 160 GB+ free disk
• Linux: Ubuntu 22.04+ or similar (x64 or arm64), NVIDIA driver installed, same VRAM and RAM guidance
• macOS: Apple Silicon, macOS 13 Ventura or newer, ≥15 GB free RAM

The 160 GB disk figure is not a typo — model weights, a Python environment, and outputs add up fast.

The Mac RAM trap nobody warns you about

This deserves its own section because it produces a support question that looks like a bug and is not.

The app checks free RAM, not total, and it checks only once — at launch, when the Python backend starts. It is never re-checked while running.

So a 32 GB Mac with a browser full of tabs open can land in API-only mode, and freeing memory afterward does nothing, because the check has already happened.

The fix from the repo is exact: close memory-heavy apps (browsers are the usual culprit), then quit and relaunch LTX Desktop. Freeing memory while the app is open has no effect.

If you have ever concluded "LTX Desktop doesn't do local generation on my Mac," this is very likely why.

What LTX Desktop can do

More than the ComfyUI templates cover, and it includes an editor:

• Text-to-video, image-to-video, audio-to-video generation
• Retake — video edit generation (regenerate a region of a clip)
• Text-to-image, and image editing (image-to-image)
• LoRA support for local video generation
• Catalog-aware prompt enhancement — locally or via the Gemini API
• A video editor interface with editing projects

The app generates locally with the LTX-2.3 (22B) model. Retake, notably, is listed as a paid API feature rather than a local one.

Neither app nor node graph appeals? If you want output without installing anything at all, ltx23.app runs LTX 2.3 in the browser. Use it to decide whether the model suits your work before committing 160 GB of disk to a local install.

The free API key — read this carefully

LTX Desktop's API relationship is more nuanced than "free tier vs paid," and getting it wrong costs either money or performance:

Free:

• Cloud text encoding — explicitly free, and the README calls it highly recommended because it speeds up inference and saves memory
• Prompt enhancement

Paid:

• API-based video generation (required on macOS and unsupported Windows hardware)
• Retake

So even in fully local mode on a capable Windows box, there is a reason to hold an API key: offloading text encoding to the cloud for free avoids downloading and running the encoder locally, which saves memory during generation.

The alternative is the Local Text Encoder, an extra download that enables fully-local operation on supported Windows hardware. Choose based on whether you care more about privacy and offline capability, or about memory headroom and speed. Keys come from the LTX Console.

LoRA handling is smarter than you would expect

This is where LTX Desktop quietly beats a hand-rolled setup.

What is supported: the app generates with LTX-2.3 (22B), and LoRAs labeled LTX-2 or LTX-2.3 both target this model. That includes LoRAs exported from ComfyUI for LTX-Video — their key names are remapped automatically.

What silently does nothing: a LoRA whose tensors do not match is skipped rather than erroring. LoRAs built for SDXL, Wan, Hunyuan, or older LTXV 0.9.x target a different architecture and will not take effect.

That skip-instead-of-crash behavior is friendly, but it means "I loaded a LoRA and nothing changed" is a real failure mode with no error message. Check the architecture before blaming strength settings.

Where files go — library downloads get their own subfolder, and your own files go directly in loras/:

models/
└── loras/
    ├── cinematic.safetensors
    ├── claymation.safetensors
    └── <lora-id>/
        └── model.safetensors

Detection rule worth memorizing: a file counts as a LoRA only if it lives in a folder named loras/lora or its filename contains lora. Many LoRAs are not named that way, so the folder is the reliable path. Subfolders are scanned recursively.

Why the LoRA selector might be missing. It appears only when all three are true:

1. You are in Video generation mode — not Image, Retake, or IC-LoRA
2. You are generating locally, not via API
3. At least one LoRA file is detected

And the installed-LoRA list is read on load, so files added while the app is running need a reopened generation view or a restart.

The built-in Browse LoRAs / Browse IC-LoRAs library is a nice touch: previews, instructions, and Hugging Face links, with entries authored by LTX marked official and everything else flagged as community-contributed and unendorsed.

Where the app data lives

Weights download into the models/ subfolder. On first launch you may be prompted to review and accept model license terms — the text is fetched from Hugging Face, so first run needs internet even for a local setup.

LTX Desktop vs ComfyUI vs the Python package

The honest positioning: LTX Desktop is the right default for most individual creators, especially on a Mac, where the alternatives get awkward fast. ComfyUI wins when you need node-level control; the Python package wins when you need pipelines the UI does not expose, like HDR output or text-to-audio.

FAQ

Does LTX 2.3 run on a Mac? Yes, through LTX Desktop on Apple Silicon with ≥15 GB free RAM, using MPS. Intel Macs are API-only.

How much VRAM do I need for local generation? The app's threshold is ≥16 GB VRAM on Windows and Linux.

Is LTX Desktop free? The app is open source. Text encoding and prompt enhancement via the API are free; API-based video generation and Retake are paid.

Why is my Mac stuck in API-only mode? The free-RAM check runs once at launch. Close memory-heavy apps, then quit and relaunch.

Can I use my ComfyUI LoRAs? Yes — LoRAs exported from ComfyUI for LTX-Video are remapped automatically. Non-matching architectures are skipped silently.

Why can't I see the LoRA selector? It requires Video mode, local generation, and at least one detected LoRA file.

Bottom line

LTX Desktop is the most under-publicized part of the LTX 2.3 ecosystem: an official, open-source app that runs the same 22B model locally, works on Apple Silicon, ships a LoRA library and video editor, and degrades to an API instead of failing on modest hardware.

If your hardware sits below its thresholds — or you just want to see output today — ltx23.app gets you there in a browser tab.

Sources

Verified July 2026 against primary documentation:

1. Lightricks/LTX-Desktop — hardware table, system requirements, RAM-check behavior, features, LoRA rules, data paths, API key policy
2. LTX Console — where the free API key is generated
3. Lightricks/LTX-2 — official GitHub repository — the underlying model and pipelines
4. Lightricks/LTX-2.3 — Hugging Face model card — the 22B model the app runs
5. ltx-core package documentation — architecture background
6. LTX-2 available pipelines documentation — Retake and other capabilities in CLI form
7. LTX-2 optimization documentation — memory considerations
8. ComfyUI official LTX-2.3 tutorial — the ComfyUI alternative
9. Lightricks/ComfyUI-LTXVideo — official ComfyUI integration
10. google/gemma-3-12b-it-qat-q4_0-unquantized — the text encoder that cloud encoding replaces
11. ltx.io — official LTX site — product hub

LTX Desktop is in Beta; thresholds, features, and pricing behavior may change between releases. Re-check the repository README before relying on any figure here.