Wan 2.2 and LTX-2 on Runpod

What are Wan 2.2 and LTX-2?

Wan 2.2 is Alibaba’s open video family. Its newer version, Wan 3.0, is sold per second through Alibaba Cloud’s API instead. Wan 2.2’s GitHub page lists a 5B model that turns text or an image into 720P video (Wan2.2 on GitHub). Its 14B models cover text to video, image to video, speech to video and character animation.

The “B” is billions of parameters, a rough measure of model size. Bigger models need more graphics memory, called VRAM, which is the deciding number when you rent a card.

LTX-2 is Lightricks’ open model, and its 2.3 version makes video and sound in one pass. The LTX-2.3 model card says it is “designed to generate synchronized video and audio within a single model”, and points ComfyUI users to the built-in LTXVideo nodes.

Ultimate LTX 2 All-In-One Workflow Guide for ComfyUI (+ Custom Audio)James Doss AI

Both run in ComfyUI, a free app where each step of the job is a box on a canvas, wired to the next. Wan 2.2 has been in ComfyUI since 28 July 2025.

Can you use Wan 2.2 and LTX-2 clips in ads?

Yes for both, with one revenue condition and a list of prohibited uses on LTX-2. Runpod rents the hardware, so the model’s own licence decides what you may do with the clips.

Wan 2.2 is under Apache 2.0, a permissive licence that allows commercial use. The Wan team’s repository states: “The models in this repository are licensed under the Apache 2.0 License.”

LTX-2 is under the LTX-2 Community License Agreement, dated 5 January 2026. It grants use “for any purpose”, subject to the prohibited uses in its Attachment A, such as deepfakes made without consent. It covers every LTX-2 version released since that date, LTX-2.3 included.

For ad work, the revenue line is the condition to check. Entities with annual revenues of at least $10,000,000 “are required to obtain a paid commercial use license” (LTX-2 licence).

The threshold counts the whole group, subsidiaries included. Below it, ad use is royalty-free, and Lightricks “claims no rights in the Output you generate”.

MiniMax H3’s open licence excludes the US, the EU, the UK and South Korea (MiniMax H3 licence). For ads there, it stays off the list.

How much VRAM do you need, and which Runpod GPU fits?

Match the model and resolution to the card’s memory first, then compare price. When a model does not fit, ComfyUI moves part of it into the pod’s ordinary memory, which works but runs much slower.

Runpod’s own Wan 2.2 guide, updated 22 October 2026, gives memory tiers by job (Runpod’s Wan 2.2 guide). The rates in the table come from Runpod’s pricing page.

The fp8 in the table is a lighter file format: the model’s numbers are stored with less precision, so it takes about half the memory.

On our own Runpod account we see the same wall from the image side. Image models of 6 to 17 GB sit comfortably on a 5090’s 32 GB. A video model of about 40 GB overflows it and offloads.

That is why a 96 GB card can cost less per clip than a cheaper card that has to offload.

What does one clip cost on Runpod?

From under a cent for a short LTX-2 clip to about a dollar for a 720p Wan 14B clip on Alibaba’s reference script. Each figure below is computed: a published generation time multiplied by Runpod’s hourly rate.

Runpod measured the 71 and 513 seconds on a warm run (models already loaded) on an RTX 4090D, the 4090’s export version with fewer cores. They are priced here at the 4090 rate.

NVIDIA’s three minutes for 8 seconds includes the slowdown when the clip outgrows 32 GB. The Wan team’s 9 minutes is an “under” figure, so that line is a ceiling. Loading models before the first clip bills too and sits outside these rows.

The Lightning LoRA, a small add-on file that cuts Wan’s sampling to 4 steps, takes the same clip from about 513 seconds to about 71.

For scale, here is a figure measured on our own account. A FLUX 3 Klein fast image on a Secure Cloud RTX 4090 takes about 28.5 seconds, or $0.007 each. The ComfyUI on Runpod guide has that table.

RTX 5090 or RTX 4090 for Wan in ComfyUI?

The RTX 4090 for the 5B model and for 14B at 480p; the RTX 5090 when 24 GB runs out. The 5090 costs about a third more an hour on Secure Cloud, so it pays only when the extra 8 GB stops ComfyUI from offloading. On Community Cloud it costs about twice as much ($0.69 against $0.34).

Runpod’s guide puts the 5090’s 32 GB at the 14B jobs that need headroom: longer clips, or LoRAs loaded on top. For LTX-2 the 5090 is the card to rent, since NVIDIA’s 4-second 720p clip fits inside its memory.

Past 32 GB, skip both. A 48 GB L40S costs less an hour than a Secure Cloud 5090. For full 14B at 720p, the 96 GB RTX PRO 6000 holds the whole model.

How to set up Wan 2.2 with HearmemanAI’s template

This route downloads the models and workflows for you on the pod’s first start. It uses a community template, published on Runpod by HearmemanAI. Attach a network volume (Runpod’s permanent disk, kept when the pod is gone) so the download happens once.

1. Find the template. In the Runpod console, open Pods, then Deploy, and search “comfyui”. Pick “One Click - ComfyUI Wan2.1 - Wan 2.2” from the community templates beside the official one.

2. Switch on the Wan 2.2 download. Open Edit Template before you deploy. Every download flag starts as false, so set download_wan22 to true, download_wan21 too if you also want the Wan 2.1 workflows, and download_wan_animate for character animation.

3. Pick a card that runs CUDA 13. Runpod’s guide says the template “is built for CUDA 13.0 and above”, CUDA being the driver layer for the GPU. Filter the picker by CUDA version. On our account, a template’s CUDA pin was one cause of “no instances available”.

4. Deploy and wait for the first boot. The pod downloads the models before ComfyUI opens. Then click Connect and open the ComfyUI link.

5. Load a Wan 2.2 workflow and run it. The template ships its workflows ready, so you set the prompt or image, the size and the clip length, then queue it.

WAN 2.2 Tutorial: Massive Upgrade to AI Video GenerationEndangered AI

6. Stop the pod when the clips are saved. It bills every second it runs, even while it sits idle.

The official ComfyUI template route

The official template gives you a clean ComfyUI and you add the Wan files yourself. It takes longer the first time and leaves you in charge of each file.

1. Deploy the official ComfyUI template from the same picker, or its Blackwell Edition on an RTX 5090. Set the container disk to 50 GB, which deployed on our account where the 150 GB default often found no card.

2. Open the template library. In ComfyUI, search the Template Library for “Wan2.2 5B”. If nothing shows, ComfyUI’s docs say your ComfyUI is out of date. The same workflows sit in the public gallery on comfy.org, searchable by model.

3. Put each file in its folder. For the 5B model, ComfyUI’s docs list wan2.2_ti2v_5B_fp16.safetensors in models/diffusion_models/, wan2.2_vae.safetensors in models/vae/ and umt5_xxl_fp8_e4m3fn_scaled.safetensors in models/text_encoders/ (ComfyUI Wan 2.2 docs).

4. Set the size and length in the Wan22ImageToVideoLatent node. For the 14B models it is the EmptyHunyuanLatentVideo node.

5. Run the queue and save the clip. Then stop the pod.

For LTX-2 the same official template works. Install the LTXVideo nodes from ComfyUI Manager, as the model card advises. Then load a text to video or image to video workflow from Template Library, then Video.

LTX-2: One Model, Many Ways To CreateComfyUI

Can you run Wan 2.2 in ComfyUI on Google Colab?

Not reliably on the free tier. Google’s FAQ lists working through a web UI as disallowed on free runtimes, and says such runtimes are often terminated (Colab FAQ). ComfyUI is a web UI. A rented pod has no such rule, and its GPU bills only while the pod runs.

Our Runpod hub sets out the other Runpod pages, and the cloud GPU price table compares these cards across GPU clouds.