How to run LTX-Video on a cloud GPU

Rent a GPU on Deploy, then open a terminal in JupyterLab or over SSH, or open the ComfyUI template. Install from the official LTX-Video repository, or follow the ComfyUI repository that README links, download the weights that file names, and run. Copy the video off the instance before you destroy it. Destroying the instance deletes its disk.

The commands below are copied from the upstream files named in each step. They were not run on a CloudGPU instance for this page. Anything this page adds around them is marked.

What the README does and does not say about memory

The LTX-Video README model table (read 2026-11-12) does not give a GB number for ltxv-13b-0.9.8-dev, ltxv-13b-0.9.8-distilled, the FP8 rows, or ltxv-2b-0.9.8-distilled. It says the 13B dev model “requires more VRAM,” the distilled 13B has “less VRAM usage,” the 2B distilled model is for “light VRAM usage,” and ltxv-2b-0.9.6 has a “lower VRAM requirement than ltxv-13b.” This page does not turn those words into a GB number.

One GB figure in that README belongs to a different file. The 0.9.7 changelog says ltxv-13b-0.9.7-distilled-lora128 “Requires only 1GB of VRAM.” That sentence is not a requirement for the 0.9.8 checkpoints. The same changelog entry also states render times and names a GPU. This page does not repeat those render times or that GPU name.

The ComfyUI repository the LTX-Video README tells you to follow is Lightricks/ComfyUI-LTXVideo. Read the same day, its prerequisites say “CUDA-compatible GPU with 32GB+ VRAM.” Its model list on that day is LTX-2.3, not the 0.9.8 filenames: ltx-2.3-22b-dev.safetensors and ltx-2.3-22b-distilled-1.1.safetensors on Lightricks/LTX-2.3. A later section says the low-VRAM loader nodes “perform the model offloading such that generation fits in 32 GB VRAM.” That 32 GB line is about the models that ComfyUI README is installing. It is not a figure the LTX-Video 0.9.8 table states.

GET /v1/public/compute-capabilities does not list GPUs. The memory column is vram_gb from gpu-prices.json, read 2026-11-12. Listings at 32 or above meet the ComfyUI README’s 32GB+ line:

Catalog name vram_gb
RTX 4080 SUPER 32G 32
V100 32G 32
RTX 5090 32
RTX 5090D 32
A100 40G 40
A800 40G 40
RTX 4090 48G 48
RTX 4090D 48G 48
A40 48
RTX A6000 48
NVIDIA RTX PRO 5000 72G 72
RTX PRO 5000 72G 72
A800 80G 80
A100 80G 80
H800 80
RTX 6000D 83
RTX PRO 6000 96G 96
RTX PRO 6000 Max-Q 96
H20 96G 96

RTX 4090, RTX 4090D, RTX 3090 and A30 are 24 in that file. V100 16G is 16. RTX 3080 Ti is 12, RTX 2080 Ti is 11, RTX 3080 is 10, RTX 2080 is 8. Those are below the 32GB+ line. Which listing you can start is on Deploy. This page does not say which listings are available.

LTX-2.5 is a different checkpoint. The AI short film guide already discusses it. This page does not copy that guide’s wording onto LTX-2.3 or onto the 0.9.8 files.

Install the LTX-Video repository

On Deploy, choose Raw GPU. The README says the codebase was tested with Python 3.10.5, CUDA 12.2, and supports PyTorch >= 2.1.2. Copied from the installation section:

git clone https://github.com/Lightricks/LTX-Video.git
cd LTX-Video

python -m venv env
source env/bin/activate
python -m pip install -e .\[inference\]

The backslashes are in the README. The optional FP8 kernels are a different repository, Lightricks/LTXVideo-Q8-Kernels. The README says to follow that repository. This page does not copy those steps.

The inference section does not print a download command. configs/ltxv-13b-0.9.8-distilled.yaml (read 2026-11-12) sets checkpoint_path to ltxv-13b-0.9.8-distilled.safetensors and spatial_upscaler_model_path to ltxv-spatial-upscaler-0.9.8.safetensors. Both files are on the model card the README links, Lightricks/LTX-Video (file list checked 2026-11-12). The same yaml also names PixArt-alpha/PixArt-XL-2-1024-MS, MiaoshouAI/Florence-2-large-PromptGen-v2.0 and unsloth/Llama-3.2-3B-Instruct. Whether inference.py downloads those by itself was not verified.

The lines below are not in the README. They use the filenames from that yaml and the Hub repo from the model card. They were not run. How to download a Hugging Face model is the hf download reference. Run pip first with download acceleration off, then turn it on for the weight download, as the download speed guide describes. cloudgpu-net is not in the LTX-Video README.

cloudgpu-net on
hf download Lightricks/LTX-Video \
  ltxv-13b-0.9.8-distilled.safetensors \
  ltxv-spatial-upscaler-0.9.8.safetensors \
  --local-dir .
cloudgpu-net off

The README’s image-to-video command, copied. PROMPT, IMAGE_PATH, HEIGHT, WIDTH, NUM_FRAMES and SEED are the README’s own placeholders:

python inference.py --prompt "PROMPT" --conditioning_media_paths IMAGE_PATH --conditioning_start_frames 0 --height HEIGHT --width WIDTH --num_frames NUM_FRAMES --seed SEED --pipeline_config configs/ltxv-13b-0.9.8-distilled.yaml

The README says resolutions should be divisible by 32 and the frame count should be a multiple of 8 plus 1, and that the model “works best on resolutions under 720 x 1280 and number of frames below 257.” It does not say where inference.py writes the file. The library snippet in the same section passes output_path="output.mp4". This page did not run it, so the CLI output path was not confirmed. Copy whatever file the run writes before you destroy the instance.

The README also links hosted demo sites. This page does not use those. It is the local install.

ComfyUI, as that repository documents it today

The LTX-Video README says: “To use our model with ComfyUI, please follow the instructions at https://github.com/Lightricks/ComfyUI-LTXVideo/.” On 2026-11-12 that page’s title still says ComfyUI-LTXVideo, and its body is the LTX-2 / LTX-2.3 install. It says LTX-2 is built into ComfyUI core. The 32GB+ VRAM line is quoted above. The same prerequisites also list a disk-space figure. This page does not repeat disk sizes.

Deploy the ComfyUI template on Deploy. How to choose a rental GPU for ComfyUI is the card page for that template. The ComfyUI-LTXVideo quick start, copied:

  1. Open ComfyUI
  2. Click the Manager button (or press Ctrl+M)
  3. Select Install Custom Nodes
  4. Search for “LTXVideo”
  5. Click Install
  6. Wait for installation to complete
  7. Restart ComfyUI

That page says required models will be downloaded on first use, and it also lists files to place by hand. The two checkpoints go in ComfyUI/models/checkpoints:

  • ltx-2.3-22b-dev.safetensors
  • ltx-2.3-22b-distilled-1.1.safetensors

The same section names spatial and temporal upscalers under models/latent_upscale_models, a distilled LoRA under models/loras, and a text encoder: download all files from the repository it links into models/text_encoders/gemma-3-12b-it-qat-q4_0-unquantized. The link on that page is https://huggingface.co/google/gemma-3-12b-it-qat-q4_0-unquantized. Other LoRA filenames are listed in that README. This page does not invent extra ones.

Its low-VRAM section, copied in substance: use the loader nodes in low_vram_loaders.py, which it says fit generation in 32 GB VRAM, and python -m main --reserve-vram 5. That command was not run. Whether the ComfyUI template already includes Manager was not checked.

For another video model with a full ComfyUI file list, see how to run Wan 2.2 and which GPU for Wan.

When it fails

CUDA out of memory. For the ComfyUI path, the published line is 32GB+ and the low-VRAM note says 32 GB. A listing below 32 does not meet that line. The 0.9.8 README table still has no GB figure of its own. Then use how to fix CUDA out of memory on a cloud GPU.

“CUDA error: no kernel image is available” on an RTX 5090. The LTX-Video README says it was tested with CUDA 12.2 and PyTorch >= 2.1.2. It does not pin a wheel. Catalog RTX 5090 and RTX 5090D are 32 GB, which meets the ComfyUI 32GB+ line and does not answer the 0.9.8 table, because that table has no GB number. If PyTorch reports no kernel image, follow the RTX 5090 guide, which says that card needs a PyTorch build for CUDA 12.8 or newer. This page did not re-test CUDA 12.2 on an RTX 5090.

Save the video, then destroy

Copy the MP4 off the instance before you destroy it (rsync or scp). Save your work before you destroy a cloud GPU is the checklist.

The pricing page states the billing rule in these words:

Per minute, rounded up, with a 1-minute minimum. One hour is pre-authorised while an instance runs, and unused minutes go back to your balance when you destroy the instance. There is no subscription and no minimum term.

For the 0.9.8 files, the model card links LTX-Video-Open-Weights-License-0.X.txt. Read that file before commercial use. This page does not restate it. LTX-2.3 has its own model card. Read the licence on the files you actually download.

Frequently asked questions

How much VRAM does LTX-Video publish?

The LTX-Video README model table does not publish a GB figure for the 0.9.8 checkpoints. It says the 13B dev model requires more VRAM, the distilled 13B uses less VRAM, and the 2B distilled model is for light VRAM usage (read 2026-11-12). A 1 GB sentence in that README applies only to ltxv-13b-0.9.7-distilled-lora128. The ComfyUI-LTXVideo README linked from there says a CUDA-compatible GPU with 32GB+ VRAM, and that its low-VRAM loaders fit generation in 32 GB VRAM. On the day this page was written, the checkpoints in that ComfyUI README were LTX-2.3 files, not the 0.9.8 files.

Which rented GPUs meet the 32 GB ComfyUI line?

vram_gb of at least 32 on https://cloudgpu.app/gpu-prices.json, read 2026-11-12: RTX 4080 SUPER 32G, V100 32G, RTX 5090, RTX 5090D, A100 40G, A800 40G, RTX 4090 48G, RTX 4090D 48G, A40, RTX A6000, NVIDIA RTX PRO 5000 72G, RTX PRO 5000 72G, A800 80G, A100 80G, H800, RTX 6000D, RTX PRO 6000 96G, RTX PRO 6000 Max-Q and H20 96G. RTX 4090, RTX 4090D and RTX 3090 are 24 GB in that file, which is below 32. That 32 GB line is the prerequisite in the ComfyUI README, not a figure the LTX-Video 0.9.8 table states. Which listing you can start is on https://cloudgpu.app/deploy. This page does not say which listings are available.

Where do the 0.9.8 weights come from?

The LTX-Video inference section does not print a download command. configs/ltxv-13b-0.9.8-distilled.yaml sets checkpoint_path to ltxv-13b-0.9.8-distilled.safetensors and spatial_upscaler_model_path to ltxv-spatial-upscaler-0.9.8.safetensors. Those files are on https://huggingface.co/Lightricks/LTX-Video (checked 2026-11-12). The ComfyUI-LTXVideo README instead names ltx-2.3-22b-dev.safetensors and ltx-2.3-22b-distilled-1.1.safetensors on https://huggingface.co/Lightricks/LTX-2.3.

Which licence applies?

Read the licence file the model card links for the checkpoint you run. For the 0.9.8 files that card links LTX-Video-Open-Weights-License-0.X.txt. This page does not restate that licence. LTX-2.3 has its own model card.

Sources (read 2026-11-12, Asia/Shanghai)

Cards you can start are on the Deploy page. Prices are on https://cloudgpu.app/pricing and are not repeated here.

Last updated: 2026-11-12 (Asia/Shanghai).