Run LTX-2.5 on a cloud GPU

The LTX-2.5 model generates video with synchronized audio from text, images and video, and its distilled BF16 files total 71.11 GB, so a 96 GB RTX PRO 6000 holds them with about 17 GB left (calculated) while a 48 GB card needs the int8 files that ComfyUI uses. Lightricks' ComfyUI README lists a GPU with 32 GB or more of VRAM as a prerequisite for its LTX-2 workflows. The weights are gated on Hugging Face with automatic approval, and the LTX-2.x Community License requires a paid license from entities with annual revenues of at least $10,000,000.

LTX-2, LTX-2.3 and LTX-2.5: how they relate

LTX-2.5 is the latest model in the LTX-2 family. The LTX-Video repository says it was released on 11 August 2026 and supersedes both LTX-Video (v0.9.x) and LTX-2. LTX-2 was created on Hugging Face on 3 January 2026 with about 18.9 billion parameters in its metadata, and LTX-2.3 followed on 4 March 2026. The LTX-2.3 and LTX-2.5 file names both say 22b, and ComfyUI's tutorial says LTX-2.5 builds on LTX-2.3 with Diffusion Fidelity Rendering, native multishot scenes, a Gemma 4 12B text encoder and an optional prompt enhancer. The LTX-2 README says the files of the two models are not interchangeable and a LoRA works only with the model it was trained on, while the LTX-2.5 model card says its testing found that the large majority of LTX-2.3 LoRAs and IC-LoRAs run on LTX-2.5 without changes. Lightricks also offers LTX-2.5 through a hosted API as ltx-2-5-fast and ltx-2-5-pro, which this page does not cover.

LTX-2.5 file sizes: 71 GB in BF16, 40 GB in int8

The sizes come from the Hugging Face file list of Lightricks/LTX-2.5, checked 12 November 2026, and GB is 10^9 bytes. The repository holds one safetensors file per component.

The distilled BF16 set that the quick start downloads is 71.11 GB (calculated), which the README calls roughly 66 GiB. The ComfyUI int8 set is 39.71 GB (calculated). The dev set comes to 80.01 GB (calculated): the dev transformer, the BF16 text encoder, both VAEs, the spatial upscaler and the distilled LoRA that the two-stage pipelines ask for.

How much GPU memory LTX-2.5 needs

Lightricks publishes no peak figure for a generation. Its system requirements page lists at least 32 GB of VRAM, 32 GB of system memory and 100 GB of free storage as the minimum, and an A100 (80 GB) or H100 with 64 GB or more of system memory as the recommended configuration. The ComfyUI-LTXVideo README lists a CUDA GPU with 32 GB or more of VRAM and 100 GB or more of free disk space as prerequisites, and says its low-VRAM loader nodes make generation fit in 32 GB. The model card suggests --quantization fp8-cast and --offload cpu for low memory, and the pipelines docs describe fp8-cast as downcasting a BF16 checkpoint on the fly for any FP8-capable GPU. The LTX model page lists a 16 GB minimum in a comparison table without naming a checkpoint or a resolution. The LTX trainer README recommends 80 GB or more for its standard LoRA config and a low-VRAM config for 32 GB. The Diffusers page gives no VRAM requirement, so the fits below are calculated from the files and leave out activations.

My pick is the RTX PRO 6000 for the command-line route, because the BF16 files leave an 80 GB card only about 2.5 GB. A 48 GB card is the route for ComfyUI with the int8 files and Lightricks' low-VRAM loaders. The RTX PRO 6000 page, the 80 GB page and the 48 GB page compare the cards.

Run LTX-2.5 in ComfyUI

ComfyUI's release notes list LTX 2.5 support in v0.32.0, published on 11 August 2026, and its LTX-2.5 tutorial lists four native workflows under Template Library, Video: text-to-video, image-to-video, first and last frame, and DFR. The ComfyUI template is built from an image that DeployGPU rebuilds, so check the ComfyUI version first, as the models guide shows. The tutorial says the files live in the gated Lightricks/LTX-2.5 repository, that you accept the license and wait for approval before downloading, and that downloads fail without access. It names these files for the text-to-video workflow, apart from an optional encoder that only the prompt enhancer uses:

The repository's folders carry the same names as ComfyUI's model folders, so pointing hf download at the models folder puts each file in place. This command combines the model card's download syntax with the tutorial's file list, and it needs a token that can read gated repositories. The download guide covers installing the hf command, and sudo docker exec deploygpu-workload ls /workspace/runpod-slim/ComfyUI/models confirms the layout first. The block below opens a shell in the container and downloads from inside it:

sudo docker exec -it deploygpu-workload bash
export HF_TOKEN=paste-your-token-here
hf download Lightricks/LTX-2.5 \
  diffusion_models/ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors \
  text_encoders/gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors \
  vae/ltx-2.5-video-vae-bf16.safetensors \
  vae/ltx-2.5-audio-vae-bf16.safetensors \
  latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors \
  --local-dir /workspace/runpod-slim/ComfyUI/models

The card says the frame count must satisfy frames % 8 == 1 and width and height must be divisible by 32. The ComfyUI-LTXVideo README adds extra nodes, such as the low-VRAM loaders, and installs through ComfyUI-Manager by searching for LTXVideo. The 100 GB disk figure is worth comparing with the Disk column of the offer, and stopping the instance deletes the disk, so the stop guide shows what to copy off. The Wan 2.2 page covers the Apache 2.0 video alternative on the same template.

Run LTX-2.5 from the command line

For scripted runs, use a plain GPU VM. The model card recommends Python 3.12 or newer, CUDA 12.7 or newer and PyTorch about 2.7, so run nvidia-smi first and check the CUDA Version it shows. These are the card's commands, with the card's example prompt:

git clone https://github.com/Lightricks/LTX-2.git
cd LTX-2
uv sync
source .venv/bin/activate

hf auth login
hf download Lightricks/LTX-2.5 \
  diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors \
  text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors \
  vae/ltx-2.5-video-vae-bf16.safetensors \
  vae/ltx-2.5-audio-vae-bf16.safetensors \
  model_patches/ltx-2.5-duration-head-bf16.safetensors \
  latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors \
  --local-dir models/ltx-2.5

uv run python -m ltx_pipelines.distilled \
  --transformer-path     models/ltx-2.5/diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors \
  --text-encoder-path    models/ltx-2.5/text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors \
  --video-vae-path       models/ltx-2.5/vae/ltx-2.5-video-vae-bf16.safetensors \
  --audio-vae-path       models/ltx-2.5/vae/ltx-2.5-audio-vae-bf16.safetensors \
  --duration-head-path   models/ltx-2.5/model_patches/ltx-2.5-duration-head-bf16.safetensors \
  --spatial-upsampler-path models/ltx-2.5/latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors \
  --prompt "A golden retriever running through a sunny meadow, cinematic lighting" \
  --seed 42 \
  --output-path output_distilled.mp4

Without --num-frames the duration head picks a length from the prompt, and --num-frames 121 sets one. The card's low-memory option is two more flags on the same command, --quantization fp8-cast and --offload cpu. It says the *-comfy-int8-convrot files are ComfyUI only and the PyTorch path needs the BF16 files. For Diffusers, the card points to Lightricks/LTX-2.5-Diffusers, which is gated too, and tells you to install Diffusers from main with pip install git+https://github.com/huggingface/diffusers. The release notes of Diffusers 0.40.0, published on 20 August 2026, list LTX-2.5 already.

The LTX-2.x Community License and the revenue threshold

The license file, LICENSE-2_x, is dated August 11, 2026 and amended October 29, 2026. Section 2.1 grants the right to use, reproduce, prepare, distribute, publicly display, publicly perform, sublicense, copy, create derivative works of, and make modifications to LTX-2.x "for any purpose, subject to the restrictions set forth in Attachment A", with one proviso: "Entities with annual revenues of at least $10,000,000 (the "Commercial Entities") are required to obtain a paid license for any use (excluding use solely for a Non-Commercial Purpose as set forth in Section 2.2) of LTX-2.x and Derivatives of LTX-2.x". It calls that paid license a Commercial Use Agreement. Section 1.6 counts subsidiaries, affiliates and companies under common control together when it checks a threshold. The model card puts the same split as "Under $10M annual revenue" and "Over $10M annual revenue", and the license says "at least $10,000,000".

Section 2.2 lets a Commercial Entity use LTX-2.x without the paid license "solely for a Non-Commercial Purpose". It says that use "for revenue-generating activity in any manner, whether direct or indirect", use "in direct interactions with or that has impact on end users", and use "to train, fine-tune, or distill any model (including any Derivative of LTX-2.x) for commercial use" are not a Non-Commercial Purpose. Section 3.5 says that a Commercial Entity receiving any Derivative of LTX-2.x, "including any fine-tuned weights, LoRA adapters, or similar adaptations", "must obtain a paid license from Licensor prior to any use of any Derivative of LTX-2.x, regardless of who created such Derivative". Section 5 says "Except as set forth herein, Licensor claims no rights in the Output you generate using LTX-2.x." Section 6 says you shall not remove, disable, alter, or circumvent any "watermarking, content provenance, latent disclosure, or other transparency features" built into LTX-2.x. The Python code in the repository is Apache 2.0, and the weights are not.