LTX-2 Local Setup: ComfyUI Install + Real VRAM Requirements

To run LTX-2 locally, plan for a 24GB GPU as the practical floor and 32GB+ for the intended experience: Lightricks' own ComfyUI extension lists a CUDA GPU with 32GB+ VRAM and 100GB+ free disk, and the checkpoint files back that up — the LTX-2 19B dev model is 43.3GB at BF16, 27.1GB at FP8, and 20GB at FP4 on disk. The fastest working setup is ComfyUI's built-in LTX-2 templates (Template Library → Video), which download the models for you and handle offloading; the current-best setup is Lightricks' ComfyUI-LTXVideo custom node running the newer LTX-2.3 22B distilled checkpoint on its fixed 8-sigma schedule. On anything under 24GB, LTX-2 is an offloading exercise, not a comfortable daily driver — alternatives at the end.

LTX-2 earned the attention it is getting: it is, per Lightricks, the first DiT-based audio-video foundation model — synchronized video and audio from a single network — and the Hugging Face numbers are absurd — the LTX-2.3 weights alone were pulled about 2.07 million times in the past month, plus another 819K for the FP8 repo and 424K for the original LTX-2 (Hugging Face, retrieved August 2026). But most of the "runs on consumer GPUs" chatter undersells what this model actually weighs. This guide gives you the real numbers first, then three verified install paths.

What LTX-2 is (and what LTX-2.3 changed) {#what-is-ltx-2}

LTX-2 is Lightricks' DiT-based audio-video foundation model — one network that generates the picture and the soundtrack together, from text, an image, or audio. Lightricks calls it the first DiT-based audio-video foundation model, and at this scale the single-model audio sync is the genuine novelty — though not an absolute first: Ovi, an open-weights 11B twin-backbone model from late 2025, generates synced video and audio at 5-second scale. The mainstream open video models you run locally — Wan, HunyuanVideo, LTX-Video 0.9.x — all generate silent clips you dub afterwards.

The short history, because the version churn matters for what you download:

Two version families coexist right now, and it trips people up. ComfyUI's built-in templates use the original LTX-2 19B checkpoints. Lightricks' own ComfyUI-LTXVideo custom node targets LTX-2.3 22B. Both work; 2.3 is where active development is (the GitHub repo — 8.5K stars — was pushed this month). If you just want output today, start with the native templates and upgrade to 2.3 when you care about the quality delta.

Both generations share the same architecture facts, verified from the model cards: a Gemma 3 12B text encoder (google/gemma-3-12b-it-qat-q4_0-unquantized), width and height must be divisible by 32, frame count must be divisible by 8 + 1, and the distilled checkpoints run at CFG 1 on a fixed short sigma schedule. The official example renders 121 frames at 24fps — about a 5-second clip — before the temporal upscaler doubles the frame rate.

LTX-2 VRAM requirements: the real numbers {#vram-requirements}

The only official hardware statement Lightricks publishes is in the ComfyUI-LTXVideo README: a CUDA-compatible GPU with 32GB+ VRAM and 100GB+ free disk. There is no official 12GB or 16GB tier — the low-VRAM stories you see are offloading, with the speed penalty that implies.

Checkpoint sizes are the honest way to reason about fit, because weights on disk set the floor for what must stream through your card. These are the published file sizes from the Lightricks/LTX-2 Hugging Face repo (retrieved August 2026):

On top of the diffusion model you still need the Gemma 3 12B text encoder and the audio/video VAEs resident somewhere, which is why the "fits fully" column is tighter than the raw file size suggests. The LTX-2.3 22B files run larger again — 46.1GB per BF16 checkpoint, with the full 2.3 repo at 157GB (Hugging Face file listing, August 2026) — hence the official 32GB+ line.

What that means per card class:

• 32GB (RTX 5090): the intended consumer target. FP8 distilled fits with headroom for activations. If you are deciding whether that card is worth street price right now, read our GPU price and memory-shortage buying map first.
• 24GB (RTX 4090 / 3090): the practical floor. FP4 (20GB) fits with little headroom; FP8 runs with ComfyUI streaming weights in and out. Generations complete, just slower per clip. See what else a 24GB card does well on our RTX 4090 local AI picks.
• 16GB and below: possible only via aggressive offload to system RAM or disk (the pipeline exposes --offload cpu and --offload disk flags). Expect long waits per clip. We honestly do not recommend LTX-2 as your first local video model on these cards — start with the lighter options in our low-VRAM text-to-video roundup instead.

Disk is the other budget line people miss: the full LTX-2 Hugging Face repo totals 314GB. Download the one quantization you will use, not the repo.

Path 1: ComfyUI native templates (fastest start) {#comfyui-native}

If you already have ComfyUI, you are five clicks from a first clip: update ComfyUI, open Template Library → Video, pick an LTX-2 workflow, and let it auto-download the models. LTX-2 has had native support since the weights dropped, so no custom nodes are needed for this path.

1. Update ComfyUI. The LTX-2 templates ship with current builds — old installs will not show them. New to the node graph entirely? Do our complete ComfyUI guide first; this page assumes you know what a loader node is.
2. Open the Template Library → Video category and pick an LTX-2 workflow. ComfyUI's official tutorial lists seven: Text-to-Video and Image-to-Video (each in standard and distilled variants), plus Depth-to-Video, Canny-to-Video, and Pose-to-Video.
3. Let the template fetch the checkpoints. The native workflows pull the 19B family — ltx-2-19b-dev (BF16), ltx-2-19b-dev-fp8, or ltx-2-19b-distilled — plus the spatial and temporal upscalers. Pick the distilled/FP8 combination unless you have datacenter VRAM.
4. Write a prompt and queue. ComfyUI's docs recommend keeping prompts within 200 words. Describe the audio you want as well as the visuals — it is one model generating both.

Start with the distilled template. A fixed 8-sigma schedule at CFG 1 versus the dev model's 40-step, CFG 4 reference settings is the difference between iterating on prompts and watching a progress bar.

Path 2: ComfyUI-LTXVideo custom node (current LTX-2.3) {#comfyui-ltx-2-3}

To run the current LTX-2.3 22B model, install Lightricks' own custom node: ComfyUI → Manager → Install Custom Nodes → search "LTXVideo" → Install. This is the repo with active development (4K stars, workflows updated for every point release) and it is where the 2.3-only features live: IC-LoRA Union Control (depth + pose + edges in one), motion tracking, HDR, lipdub, and a text-to-audio workflow.

After installing the node, download the 2.3 files into these exact folders — locations verified from the ComfyUI-LTXVideo README:

Then load one of the bundled example workflows — the two-stage distilled text/image-to-video is the sensible default. Two-stage means the model generates at a lower base resolution and the spatial upscaler doubles it, which is how LTX-2 keeps generation practical: you spend your VRAM on the first pass and the 996MB upscaler handles the resolution jump.

There are FP8 repos for 2.3 on Hugging Face as well (LTX-2.3-fp8, with dev and distilled builds, plus an LTX-2.3-nvfp4 repo whose distilled variant was still marked "coming soon" as of August 2026). Same logic as the 19B family: take the smallest quantization that keeps quality acceptable, because at these file sizes every gigabyte is fit-or-stream.

Path 3: the CLI pipeline (no ComfyUI) {#cli-setup}

If you want scripted, repeatable generations, Lightricks ships a Python pipeline in the LTX-2 GitHub repo. Commands below are verified against the current README (the repo uses uv; Lightricks' model cards list CUDA 12.7+ and PyTorch ~2.7):

# 1. Clone the repo
git clone https://github.com/Lightricks/LTX-2.git
cd LTX-2

# 2. Authenticate with Hugging Face and download the checkpoints
hf auth login
# then use "hf download" per the README to fetch the ltx-2.3 checkpoints
# and the Gemma 3 text encoder into models/

# 3. Generate with the distilled pipeline
uv run python -m ltx_pipelines.distilled \
    --distilled-checkpoint-path models/ltx-2.3/ltx-2.3-22b-distilled-1.1.safetensors \
    --spatial-upsampler-path models/ltx-2.3/ltx-2.3-spatial-upscaler-x2-1.1.safetensors \
    --gemma-root models/gemma-3-12b \
    --seed 42 \
    --output-path output.mp4 \
    --prompt "your scene description, including the audio you want"

The flags that matter when VRAM is tight, straight from the repo docs:

• --quantization fp8-cast (for BF16 checkpoints) or fp8-scaled-mm (for FP8 checkpoints on Hopper+) — run the transformer in FP8 to cut memory, per the pipeline docs.
• --offload cpu or --offload disk — spill weights out of VRAM. This is the mechanism that makes sub-24GB cards technically possible and practically slow.
• On Hopper and newer datacenter GPUs the repo recommends FlashAttention 3 (FlashAttention 4 on B200) — irrelevant for consumer cards, useful if you rent.

The distilled pipeline is the one to use here too — "fastest inference with 8 predefined sigmas" (8 steps in stage 1, 4 in stage 2), per the repo. The pipeline also exposes --enhance-prompt, which rewrites your prompt with a built-in enhancer before generation — worth trying when a scene comes out flat.

Settings and hard rules that will save you an evening {#settings}

Three constraints are enforced by the model, not suggestions: width and height must be divisible by 32, frame count must be divisible by 8 + 1 (so 121, not 120), and the distilled checkpoints expect CFG 1 on their fixed sigma schedule. Feed it 120 frames or CFG 7 out of habit from image models and you get errors or mush. For calibration, the dev checkpoint's model-card reference example runs 121 frames at 768×512, 40 steps, CFG 4; the distilled two-stage example drops to CFG 1 with as few as 3 steps in its second stage.

Practical defaults that follow from the official examples and docs:

• Start at 121 frames, 24fps (~5 seconds), then let the temporal upscaler take you to higher fps rather than generating more frames.
• Prompt the audio explicitly. "Rain on a tin roof, distant thunder" produces a soundtrack; silence in the prompt leaves the audio model guessing.
• Keep prompts under ~200 words (ComfyUI's stated guidance).
• Use two-stage workflows for resolution. Generate low, upscale x2 spatially. Lightricks' 4K/50fps capability claims assume the full multi-stage pipeline, not a single monster pass.

Honest limitations {#limitations}

LTX-2 is the most capable open video model you can run at home, and it is also heavier, hungrier, and younger than its marketing. Know these before you clear 100GB of disk:

• The VRAM story is 32GB-native, 24GB-workable, below-that-painful. No amount of enthusiasm changes a 27.1GB FP8 file. Lightricks' own extension README says 32GB+; everything below that is offloading.
• Non-speech audio is the weak spot — by Lightricks' own admission. The model card states: "When generating audio without speech, the audio may be of lower quality." Dialogue and lip-sync (there is a dedicated lipdub workflow) are the showcase; ambient soundscapes are hit-and-miss.
• The license is not Apache. The weights ship under the LTX-2 Community License Agreement. It permits derivatives under its terms, but if you are generating for commercial work, read the actual agreement on the model card before you build a pipeline on it — do not assume standard open-source terms.
• Version churn is real. 2.0 to 2.3 took roughly three months, checkpoints get -1.1 refreshes, and workflows change with them. Expect to re-download multi-gigabyte files a few times a year.
• Clips are short. The official examples render 121 frames at 24fps — about five seconds — and the temporal upscaler doubles frame rate, not duration. For longer silent footage, LTX-Video 0.9.x and Wan-family models stretch further — see our Wan video generation guide.

Under 24GB? Run these instead {#alternatives}

If your card has 16GB or less, spend your time on models that fit it natively rather than fighting LTX-2's offload flags. Our low-VRAM text-to-video guide covers the current options tier by tier, HunyuanVideo remains the strongest silent-video alternative in the 12-24GB band, and the local AI video generation pillar compares LTX-2, Wan, and Hunyuan head-to-head so you can pick by hardware instead of hype. You lose the synchronized audio — among big open video models that is LTX-2's party trick, though the smaller open-weights Ovi (11B) also does synced audio-video at 5-second scale — but you gain a model that actually fits your card.

Key takeaways {#key-takeaways}

1. Real requirements: 32GB+ VRAM official, 24GB practical floor, 100GB+ disk. File sizes tell the truth: 43.3GB BF16, 27.1GB FP8, 20GB FP4 for the 19B family; LTX-2.3 22B is larger still.
2. Fastest start: ComfyUI Template Library → Video → LTX-2 distilled. Auto-downloads, no custom nodes, 19B checkpoints.
3. Best current model: LTX-2.3 22B via the ComfyUI-LTXVideo custom node (install through ComfyUI Manager, search "LTXVideo") — distilled-1.1 checkpoint, two-stage upscaling, IC-LoRA controls.
4. Respect the hard rules: dimensions divisible by 32, frames divisible by 8 + 1, distilled = CFG 1 on its fixed sigma schedule, prompt the audio explicitly.
5. Under 24GB, run Hunyuan or Wan instead — you trade away synchronized audio for a model that fits.

FAQ {#faq}

How much VRAM does LTX-2 actually need?

Lightricks' ComfyUI extension officially lists a CUDA GPU with 32GB+ VRAM and 100GB+ free disk. In practice a 32GB RTX 5090 runs FP8 (27.1GB on disk) comfortably, and a 24GB RTX 4090 or 3090 runs FP4 (20GB) or FP8 with weight offloading at reduced speed. There is no official 12GB or 16GB tier — anything below 24GB relies on CPU/disk offload and is slow enough that we point those cards at lighter models instead.

Can I run LTX-2 on a 12GB or 16GB GPU?

Technically the pipeline supports it — the CLI exposes --quantization fp8-cast and --offload cpu / --offload disk, and ComfyUI streams weights automatically. Practically, you are moving 20GB+ of weights through a card that holds a fraction of them, per generation. It works as a demo, not a workflow. On these cards we recommend HunyuanVideo or a Wan-family model sized for your VRAM.

Does LTX-2 really generate audio and video together?

Yes — that is its defining feature. LTX-2 is a DiT-based audio-video foundation model that generates synchronized video and sound in a single model, including dialogue with lip-sync (LTX-2.3 ships a dedicated lipdub workflow). One documented caveat from the model card: audio without speech "may be of lower quality," so ambient-only soundscapes are less reliable than scenes with dialogue.

LTX-2 vs LTX-2.3 — which should I download?

LTX-2.3 (22B), if your setup allows it — Lightricks calls it "a significant update" with improved audio, visual quality, and prompt adherence, and it is where new workflows land. The catch: ComfyUI's built-in templates still use the original 19B checkpoints, which are the easiest zero-config start. Reasonable path: first clip on the native 19B distilled template, then install the ComfyUI-LTXVideo custom node for 2.3.

Is LTX-2 free for commercial use?

The weights are free to download and self-host under the LTX-2 Community License Agreement — a custom license, not Apache or MIT. It allows model derivatives under stated terms, but commercial-use conditions are defined by that agreement, so read it on the Hugging Face model card before shipping paid work. This is the same pattern as most recent open-weights video models: free for local use, custom terms beyond that.

Sources {#sources}

• Lightricks/LTX-2 on GitHub — pipeline code, quickstart commands, quantization/offload flags
• Lightricks/ComfyUI-LTXVideo — official custom node, 32GB+/100GB+ requirements, folder locations, LTX-2.3 workflows
• LTX-2 model card on Hugging Face and the LTX-2.3 model card — checkpoint list and sizes, dimension/frame rules, audio caveat, license
• ComfyUI docs: LTX-2 native workflows — template names and prompt guidance
• Lightricks LTX-2 announcement (October 2025) — native-4K / 50fps / synchronized-audio capability claims, quoted as vendor claims