Updated September 2026 · New recommended cards (RTX 50 series) and an updated table showing how much video memory today's models need. Part of our AI and ComfyUI tutorials (in Italian).
When we started running AI on our own computer, our first question was the same one you keep asking us: "is my graphics card good enough?". Today we work with ComfyUI on a perfectly ordinary RTX 3060 12 GB, and we can tell you something you won't find on any spec sheet: video memory matters more than raw power.
VRAM decides which models you can run and at what resolution, and it's the reason a card from a few years ago with 12 GB can be more useful than a brand-new one with 8. In this guide we explain how much video memory you need, why NVIDIA is the smart choice today and which cards we recommend in 2026, from the cheapest to the most powerful, including the 3060 12 GB we use every day.
In short: the graphics cards we recommend for AI
- To spend as little as possible: RTX 3060 12 GB, back on sale: the card we run all our tests on, perfect for SDXL and for the lighter versions of Flux.
- The best choice to get started: RTX 5060 Ti with 16 GB. Be careful not to mix it up with the 8 GB version, which is too limited for AI.
- For people who use AI every day: RTX 5070 Ti 16 GB, much faster with the same amount of memory.
- No compromises: RTX 5090 with 32 GB, for the largest video models and for training your own LoRAs.
- If you need to buy the whole computer: a ready-made PC with an RTX 5060 Ti 16 GB and 32 GB of RAM.
How much video memory do you need for AI?
For running AI locally, video memory (VRAM) matters more than power: the model has to fit into the card's memory. If it doesn't, generation slows to a crawl or stops with an out of memory error. A slower card with more memory is often a better choice than a faster one with little memory.
| Video memory | What you can do | Verdict |
|---|---|---|
| 8 GB | Stable Diffusion 1.5 and SDXL; Flux only in reduced versions and with patience; video models are almost out of reach | Just for trying things out |
| 12 GB | SDXL with no trouble, Flux in reduced versions (fp8 or GGUF), short videos with the lightest models | The recommended minimum |
| 16 GB | Flux and recent image models comfortably, video with models like Wan and LTX-Video in optimized versions | The ideal choice |
| 24–32 GB | The largest video models at full quality, high resolutions, LoRA training | For professionals |
Besides video memory, your system RAM matters too: with 32 GB, ComfyUI can offload part of the model from the graphics card to RAM when needed; with 16 GB you hit the limit quickly.
Why an NVIDIA card
Almost all generative AI software and models are developed and optimized first and foremost for NVIDIA cards, thanks to the CUDA platform and Tensor Cores, compute units designed specifically for neural networks. The result is maximum compatibility: when a new model comes out, it works on NVIDIA right away.
ComfyUI now also has packages for AMD and Intel cards, and the Desktop app runs on Macs with Apple silicon: you can use them, but with less compatibility and longer generation times. To give you an idea of the difference, here are the times we measured with ComfyUI on an RTX 3060 12 GB, compared with computers without an NVIDIA card:
IMAGE GENERATION TIMES:
ComfyUI - DreamShaper XL model (SDXL Lightning)
- 3 seconds on an RTX 3060 12 GB
- 1 and a half minutes on other GPUs (AMD, Intel, M4)
VIDEO GENERATION TIMES:
ComfyUI - LTX-Video 0.9.1
- 1 minute 20 seconds on an RTX 3060 12 GB
- over 12 minutes on other GPUs (AMD, Intel, M4)
It's no coincidence that, as we showed in our infographic on NVIDIA cards with 24 GB of memory (in Italian, based on Italian users), cards with lots of video memory are still uncommon: for most people, 16 GB is the sweet spot.
The best graphics cards for AI in 2026
1. NVIDIA RTX 3060 12 GB: the cheapest option
This is the card we used for our tests, our videos and our free ComfyUI workflows, and it's back on sale. It's the least expensive way to get 12 GB of video memory: SDXL runs with no trouble, Flux works in reduced versions (fp8 or GGUF), and so do short videos with lightweight models. It isn't as fast as newer cards, but for learning and non-professional use it's an excellent starting point. If your budget allows, the 16 GB of the RTX 5060 Ti will give you more headroom for the future.
2. NVIDIA RTX 5060 Ti 16 GB: the best choice to get started
It's the cheapest way to get 16 GB of video memory on a new card. It's not the fastest, but it runs Flux, SDXL and optimized video models without the memory problems of 8 and 12 GB cards, and it uses little power. Make sure you pick the 16 GB version: there's also an 8 GB RTX 5060 Ti that we don't recommend for AI.
3. NVIDIA RTX 5070 Ti 16 GB: the perfect balance
The same amount of memory, but far more compute power: images arrive in a fraction of the time, and video becomes genuinely usable for everyday work. This is the card we recommend to anyone who uses AI professionally but doesn't want to pay flagship prices.
4. NVIDIA RTX 5090 32 GB: the very best
With 32 GB of memory it runs even the largest video models at full quality, generates at high resolutions and lets you train your own LoRAs in reasonable time. It's expensive and needs a powerful power supply and a roomy case: it makes sense if AI is your job.
How to choose a graphics card for AI
- Video memory first: 16 GB if you can, 12 GB at the very least.
- Then speed: with the same memory, a higher-tier card generates faster.
- Power supply and space: check the manufacturer's recommended wattage and the card's length against your case.
- The rest of the PC: at least 32 GB of RAM and a large SSD, because models take up tens of GB.
If you want to compare the performance of different cards, you'll find detailed benchmarks in this Tom's Hardware article. If instead you need to buy a whole computer, read our guide to the best PC for 3D and rendering, which applies to AI as well.
Alternatives: cloud and online platforms
If you don't want to buy a new graphics card, you can use AI online. Generation platforms are the simplest option and often give you free credits to start: you'll find the main ones in our list of AI image generators. Alternatively, you can rent a graphics card in the cloud and run ComfyUI on it, paying by the hour: it's worth it for occasional use, but it takes a bit more technical know-how.
Got your graphics card? Start here
- Install ComfyUI on Windows, with 3 free workflows
- Install Stable Diffusion on your PC
- Stable Diffusion vs Flux: the models compared
Frequently asked questions
What is the best graphics card for ComfyUI?
To spend as little as possible, the RTX 3060 12 GB; for most people, an NVIDIA card with 16 GB of memory: the RTX 5060 Ti 16 GB as the balanced choice, the RTX 5070 Ti if you want more speed. The RTX 5090 32 GB is for those working with the largest video models.
Is 8 GB of VRAM enough for Stable Diffusion?
For Stable Diffusion 1.5 and SDXL, yes, but you're at the limit: newer models and video models require reduced versions and a lot of patience. If you're buying today, get at least 12 GB, ideally 16.
Can you use AI with an AMD card?
Yes: ComfyUI has a package dedicated to AMD cards. However, compatibility with new models and performance are generally lower than on NVIDIA, and if something goes wrong you'll find less help in the community.
Is it worth buying a used card?
It can be, to get more memory for the same money, for example an RTX 3060 12 GB or an RTX 3090 24 GB. Only buy from reliable sellers who offer a warranty, and check that the card hasn't been used for mining.
