Get Started
← Own your AI · Mission 25 of 31

Make AI video clips with Wan 2.2 on a rented GPU

Open video models now make five-second 720p clips that look like stock footage, and their licences let you use what you make. They need a big graphics card, so rent one by the hour: ComfyUI and Wan 2.2 on a pod, driven from your laptop through an SSH tunnel, with a batch script that queues a night of prompts and a calculator that tells you what each usable clip really cost.

150 minAdvancedTools: Roughly $1 to $3 of GPU time per session, plus a few dollars a month to keep the models on a volumeStep 1 free
ComfyUI on a rented GPU, installed by one script, reachable only through your own SSH tunnel
Three ready workflowsWan 2.2 5B text-to-video and image-to-video at 720p, and Wan 2.2 14B image-to-video in 4 steps
A batch queue that runs a list of prompts from your laptop and downloads every clip with a log
  • Free upscaling to 1080p and frame smoothing with ffmpeg on your own PC
  • A cost-per-usable-clip calculator, so you know what a keeper costs before you scale up
Before you start
  • A RunPod (or similar) account with an SSH key and a spending limit, set up in Rent a cloud GPU. Do that mission first
  • About $10 of credit. One session of experiments uses $1 to $3
  • Python via uv on your laptop for the batch script: docs.astral.sh/uv ↗
  • ffmpeg on your laptop for the upscaling step: FFmpeg for creators
  • No graphics card needed on your side. Everything heavy runs on the pod
01/08Pick the model: Wan 2.2 5B, 14B or LTX0 done
STEP 01 of 08~10 min

Pick the model: Wan 2.2 5B, 14B or LTX

Choose by what you want to make, the card you'll rent and the licence you need.

  1. Wan 2.2 from Alibaba's Wan team is the open video model most ComfyUI users reach for, and it's Apache 2.0 licensed on its model card ↗: the authors claim no rights over what you generate. Wan 2.2 is the newest version with Apache 2.0 weights on the official Wan-AI Hugging Face page ↗; later versions are offered through Alibaba's own services. Two flavours matter:

    • Wan 2.2 TI2V 5B: text-to-video *and* image-to-video in one model, 720p (1280x704) at 24fps. Wan's own card says it makes a 5-second 720p clip "in under 9 minutes on a single consumer-grade GPU" such as an RTX 4090. Start here.
    • Wan 2.2 I2V 14B: image-to-video only, noticeably better motion and detail. It's two 14B models (one for early "high noise" steps, one for late "low noise" steps). ComfyUI's template runs fp8 versions with the lightx2v 4-step LoRAs, so a clip takes 4 sampling steps instead of 20 or more.

  2. LTX-2.x from Lightricks is the other strong open family: up to 22B parameters, and it generates sound with the video in one pass. Its licence is not Apache: the LTX-2.x Community License requires a paid licence for companies with $10,000,000 or more in annual revenue (licence text ↗). Free for individuals and small businesses, but read it before client work. ComfyUI has LTX-2 templates too; this mission's kit is built for Wan.

  3. What a clip isFive seconds, one shot, one main action. Think stock footage, B-roll, a product spinning, a photo coming to life. Stories are made by editing several clips together, not by one long prompt.

  4. Get the kitvideo-kit.zip has the pod setup script, three workflows, the batch queue, example prompts and the cost calculator. kit.md has the session checklist.

✓
CheckpointYou've picked 5B (start here) or 14B image-to-video, and you know which licence covers what you'll make.
Stuck? Ask the community
Next mission · Own your AITurn invoices, receipts and forms into clean JSON with a local AI
→