Local AI software
ComfyUI, answered: the GPU you need for Flux, Stable Diffusion and Wan 2.2 video
ComfyUI is a free, open-source app for running image and video models such as Stable Diffusion, Flux and Wan on your own GPU, built from connected nodes. What you can run comes down to your graphics card's memory (VRAM). The table lists the Wan 2.2 video setups each VRAM size handles, from ComfyUI's and Runpod's own figures, with the cheapest card to rent for each, and below it is every guide we have on image and video generation, from a first install to AMD and Intel cards.
Official site: comfy.org · checked 2026-11-07
What runs on your GPU
Alibaba's Wan 2.2 is the open video model most people run in ComfyUI. What each VRAM size handles, from the sources named, and the cheapest NVIDIA card with that much memory to rent right now.
| VRAM | What it runs | Cheapest card to rent | Rent |
|---|---|---|---|
| 8 GB+ | Wan 2.2 5B, with ComfyUI's native offloading (ComfyUI's Wan 2.2 docs) | RTX 3060 from $0.033/hr | Vast.ai |
| 24 GB+ | Wan 2.2 14B at 640x640 with fp8 weights (Runpod's Wan 2.2 guide) | L4 from $0.045/hr | Vast.ai |
| 48 GB+ | Wan 2.2 14B in fp16, or in fp8 at 720p (Runpod's Wan 2.2 guide) | RTX A6000 from $0.107/hr | Vast.ai |
| 80 GB+ | Wan 2.2 14B at 720p without offloading (Runpod's Wan 2.2 guide) | RTX PRO 6000 Blackwell Max-Q Workstation Edition from $0.443/hr | Vast.ai |
For image models (Flux, SD 3.5, SDXL) see how much VRAM each needs; the full breakdown with cost per clip is in our ComfyUI cloud GPU guide. Prices are the lowest in our hourly snapshot (1h ago); rent buttons are referral links.
Install it and make your first images
- How do I run Stable Diffusion on my own PC for free?Generate unlimited AI images on your own computer — no Midjourney subscription, no per-image fees, nothing sent to a server. Here's exactly how to set it up, and what graphics card you need.
- How do I run Flux locally?Flux makes the best AI images you can generate at home, and you can run it on a normal GPU — even a 12GB card, thanks to quantized versions. Here's exactly how, and what VRAM you need.
- No NVIDIA card: Stable Diffusion and Flux on AMD and Intel ArcYour Radeon or Arc card can make AI images. The trick is picking the right backend, because three of the seven options are dead ends. We read ComfyUI's code, AMD's ROCm docs and Intel's PyTorch guide to sort them, then matched FLUX, SD 3.5 and SDXL to each card size.
Which GPU, and how much VRAM
- How much VRAM do Flux, SD 3.5 and SDXL need?SDXL fits any 8 GB card. FLUX.1 dev needs 16 GB at fp8 or 12 GB as a 4-bit GGUF, SD 3.5 Large 12 GB at fp8, and FLUX.2 dev 24 GB even at 4-bit. Every number comes from the model files on Hugging Face, re-read Nov 3, 2026, with the rule we used printed.
- Which GPU should I buy for Stable Diffusion and Flux?For most people in November 2026 it's an RTX 5060 Ti 16GB: it runs SDXL and FLUX.1-dev in ComfyUI for about $549. Below that a 12GB RTX 3060 still works, a used 24GB RTX 3090 holds all of Flux in memory, and the 32GB RTX 5090 is the fast, scarce top end. Prices and timings dated and sourced.
- What's the cheapest cloud GPU for ComfyUI and Wan 2.2?A 24 GB card runs Flux and Wan 2.2 14B video at 640x640, and on Nov 3, 2026 the cheapest one was an RTX 3090 on Vast from about $0.10 an hour. Which card each workflow needs, what one Wan clip costs in GPU time, when you really need 48 GB, and the setup that doesn't waste your first hour.
AI video: run it yourself or use an app
- An app, an API or a rented GPU: what one clip costsI read the Reddit threads, the licences and the price pages so you don't have to. What one 5-second clip really costs in Higgsfield, through an API, and on a rented RTX 4090, plus the MiniMax H3 licence line almost nobody mentions.
- Is Wan 3 free, and can I download it?Alibaba's Wan 3.0 has no open weights, so you can't run it free on your own PC. The official free routes are daily check-in credits on wan.video and a 30-second API quota. What one clip costs, what GPU the open Wan 2.2 needs instead, and why most "free Wan 3.0" sites aren't Alibaba's.
- Hosted video models (Seedance, Veo, Wan 3.0) priced per clipSeedance 2.5, Google's Veo 3.1, MiniMax's Hailuo and H3, Wan 3.0, Grok Imagine and Runway, matched to the reason you're leaving Kling. What each one does best, where to use it, what a clip costs, its free option, and what happened when we gave Kling 4.0 and Seedance 2.5 the same prompt.
Get the GPU