Knowledge Base · Local AI

How much VRAM for Stable Diffusion & image AI?

By Tiny · how I pick

Updated October 2026. No affiliate links on this page. Image models use memory very differently from text models, so here's the number for each one and where compression helps.

It depends on the model:

If you're buying a card mainly for image AI, treat 12 GB as the minimum and 16 GB as the sweet spot.

Words you'll see in this guide
  • VRAM: the memory on your graphics card. AI models load into it, so more VRAM means you can run bigger models.
  • Quantization (FP8, GGUF): shrinking a model by storing its numbers less precisely, so it fits in less memory with a small drop in quality. Think of saving a photo as a smaller JPEG. "4-bit" (Q4) is the common setting.
  • FP16: full precision, the uncompressed version of a model. Best quality, most memory.
  • LoRA: a small add-on file that teaches an image model a style or subject. Each one uses a bit more memory.
  • ControlNet: an add-on that lets you guide an image's layout or pose with a sketch or reference picture.

See the full glossary

VRAM by model

ModelVRAM neededNotes
Stable Diffusion 1.5About 4 GBRuns on almost anything, and it's still popular because it's fast
SDXLAbout 8 GB12 GB is comfortable with LoRAs or ControlNet
SD 3.5 MediumAbout 8 GBFits an 8 GB card
SD 3.5 LargeAbout 18 GB at FP16, about 11 GB at FP8Use a compressed version on regular cards
Flux.1About 24 to 33 GB at FP16, about 12 to 13 GB at FP8FP8 or GGUF versions fit 12 to 16 GB cards

Numbers include the model plus typical working memory at 1024 pixels. Higher resolutions, larger batches, and ControlNets use more.

The two groups that matter

Image AI in 2026 splits neatly into two groups.

Fits on 8 GB

SD 1.5, SDXL, SD 3.5 Medium

These run well on an 8 GB card, and SD 1.5 even works on 6 GB. If mainstream image generation is all you want, an 8 to 12 GB card covers it with room to spare, and it's fast.

Needs a compressed version

Flux, SD 3.5 Large

At full precision these need 18 to 33 GB, more than any mainstream card has. The fix is a compressed version. FP8 and GGUF builds cut the memory roughly in half, which brings Flux down to about 12 to 13 GB so it fits a 12 to 16 GB card. You give up a little speed and a little quality, and in return you can run it at all.

Quick guide:
  • 8 GB: SD 1.5 and SDXL, comfortably.
  • 12 GB: the workable middle, if you're fine using compressed versions of the newer models.
  • 16 GB: Flux and SD 3.5 Large without fuss.

Why image AI needs a graphics card more than chat does

Image models are smaller than the big text models, but each picture takes far more number crunching. That makes a separate graphics card matter more here than for chat. You can generate images on a processor alone, but it's painfully slow. For text models it's a different story: see running local AI without a GPU.

Where to go next

Quick answers

How much VRAM do you need for Stable Diffusion?

It depends on the model. SD 1.5 runs on as little as 4GB, SDXL wants about 8GB (12GB is comfortable with add-ons like LoRAs and ControlNet), and newer models like SD 3.5 Large and Flux need 12 to 16GB once compressed. 12GB is a good minimum and 16GB is the sweet spot.

How much VRAM does Flux need?

At full precision, Flux.1 needs roughly 24 to 33GB. Compressed FP8 and GGUF versions bring it down to about 12 to 13GB, so it runs on a 12 to 16GB card with a small speed and quality trade-off.

Can I run image AI on 8GB VRAM?

Yes. SD 1.5, SDXL, and SD 3.5 Medium fit on an 8GB card. Flux and SD 3.5 Large need heavy compression to fit, so 12GB or more is better if you want newer models.

Is image generation more VRAM-hungry than running LLMs?

Not bigger, but different. Image models are smaller than big chat models, and they depend much more on the graphics card for speed. Generating images on a processor is possible but very slow.

Related guides