Updated October 2026
Hardware for local image generation: Flux, SDXL and video
Image models want VRAM too, but on a different curve: 16GB is comfortable, speed comes from compute rather than bandwidth, and NVIDIA's lead is wider than it is for language models. The cards that fit, and why a Mac is the wrong tool here.
As an Amazon Associate I earn from qualifying purchases. Product links on this page go to Amazon; they are affiliate links, and the price you pay is the same.
The picks
NVIDIA GeForce RTX 5060 Ti (16GB)
SDXL at full quality and Flux at FP8. The 16GB matters; the 8GB version of this card is a dead end for Flux.
For local LLMs, holding the model matters more than speed. This is the lowest price of entry to 16GB with CUDA. Get the 16GB version; the 8GB one is useless for this.
NVIDIA GeForce RTX 5070 Ti (16GB)
The same 16GB with nearly double the memory speed and a lot more compute: roughly twice the images per minute of the 5060 Ti.
The best new 16GB card you can buy under the budget line while 24GB cards sit above it. CUDA means every tool just works.
NVIDIA GeForce RTX 4090 (24GB, used or renewed)
24GB runs Flux at FP16 with LoRAs and ControlNets loaded, and it's still the reference card every ComfyUI workflow was built on.
Mature drivers, huge community, and 24GB covers most of what people actually run. Buy renewed with a warranty; the price has held because local AI buyers want exactly this card.
NVIDIA GeForce RTX 5090 (32GB)
32GB, Blackwell's FP4 support (NVIDIA's FP4 Flux runs in 10GB), and the compute to make video generation tolerable. The only card that makes Wan and Hunyuan video feel like a tool.
Memory bandwidth near 1.8TB/s. Tokens per second on a 32B model are in a different league from anything below it.
Why image generation is a different question
Language models are bandwidth-bound: the whole model streams through memory for every token. Diffusion models are compute-bound: the GPU does the same heavy math over an image twenty or thirty times per picture. So the ranking changes. Memory still has to be big enough to hold the model, but once it is, cores and tensor throughput decide images per minute, and that’s where NVIDIA’s lead over Apple and AMD is widest.
How much VRAM the current models need
| Model | Weights | Comfortable VRAM | Notes |
|---|---|---|---|
| Stable Diffusion 1.5 | 2GB | 6GB | old, fast, still the base of many LoRAs |
| SDXL | 6.5GB | 8 to 12GB | the workhorse; refiner and ControlNets add a few GB each |
| Flux.1 dev, FP8 | 12GB | 16GB | the quality default in 2026; FP8 loses almost nothing |
| Flux.1 dev, FP16 | 23GB | 24GB | for training LoRAs and the last few percent |
| Flux.1 dev, FP4 (Blackwell) | 6GB | 12GB | RTX 50-series only; NVIDIA’s own quantization |
| Wan 2.x, Hunyuan video | 14 to 26GB | 24GB minimum | 32GB and patience; video is still slow on anything |
Text encoders (T5 for Flux) and VAEs want several more GB alongside the model, which is why 16GB is the comfortable floor and 12GB is where people start fighting their tools.
The picks, by what you’ll do
Just making images, SDXL and Flux FP8. A 16GB card. The 5060 Ti 16GB is the cheapest route; the 5070 Ti does the same work about twice as fast. Both run Flux at FP8 with a LoRA or two loaded. Neither trains.
Training LoRAs, stacking ControlNets, Flux at FP16. 24GB. The used RTX 4090 is the card: every ComfyUI workflow, every training script and every memory-saving trick was written against it first. The 3090 does the same jobs at about 60% of the speed and is the right call if the budget stops there.
Video, or you’re impatient. The RTX 5090. 32GB holds the video models without offloading, FP4 support halves memory on Flux with little visible loss, and the compute roughly doubles the 4090. Video models still take minutes per clip; the 5090 is the difference between minutes and a quarter of an hour.
What about a Mac
It works, and it’s the wrong tool. Draw Things (App Store) and ComfyUI both run on Apple Silicon, and a 128GB Mac Studio holds any image model you want. But diffusion is compute-bound, and an M5 Max does the math at a fraction of a 4090’s rate: a Flux image that takes a 4090 about 10 seconds takes a Mac Studio closer to a minute. If you already own the Mac, use it. Don’t buy a Mac for image generation.
What about AMD
Worse than for language models. ROCm on Linux runs ComfyUI and gets decent SDXL speeds on a 7900 XTX, but Flux support, FP8 and the memory-saving nodes lag NVIDIA by months, and Windows is rougher still. Budget cards from AMD’s 9000 series have 16GB and good gaming value; for diffusion, the time you lose to driver and node issues costs more than the card saves.
Software
ComfyUI is the standard: node-based, runs everything, every new model lands there first. Steep for an afternoon, then it’s the only thing you use. Fooocus and Draw Things (Mac) are the simple front ends if you want prompts in, pictures out. All of them use CUDA on NVIDIA with no setup.
Power and heat
Image generation holds the GPU at 100% for the whole run, unlike chat, which spikes and idles. A 4090 or 5090 pulls its full rated power for minutes at a time. The build guide has the power supply and case picks; for this workload, don’t shave them.