Local image generation models: SDXL, Flux and how much VRAM

For image generation at home, SDXL is the easy, well-supported choice on 8 to 12 GB cards, and Flux gives better images and better text in pictures if you have 16 to 24 GB. Older Stable Diffusion 1.5 still runs on almost anything. To try it with no install at all, our Studio generates images in your browser on your own GPU.

The main model families

The VRAM figures are approximate. They depend on image size, the tool, and its memory-saving settings. Many tools can run bigger models on smaller cards by moving parts of the model to system memory, at a cost in speed. Quantized versions of Flux, including GGUF files, cut memory a lot for a small loss in quality. See quantization explained.

Images are different from text

Text generation is limited by memory bandwidth; see why bandwidth sets speed. Image generation runs the model many times per picture and leans much harder on raw compute. So:

• Newer NVIDIA cards win clearly. Tensor cores matter here.
• A card that's great for chat can be poor for images. Our rack's Tesla P100s are fast at text but slow at images, because they have no tensor cores. See the Tesla P100 for local LLMs.
• Macs work, more slowly than a similar-priced NVIDIA card.
• AMD and Intel cards work with some tools, with more setup. See AMD GPUs for local AI.

The RTX 3060 12 GB is a popular starter for images: enough memory for SDXL with room to spare. A used RTX 3090 runs Flux comfortably.

Try it in your browser first

Our Studio runs SD-Turbo entirely on your own GPU through WebGPU. Nothing is sent to a server. It's about a 2.5 GB download the first time, then it's cached and works offline. It makes 512×512 images in a single step, so it's fast but rough next to SDXL or Flux. It's a quick way to see if your machine handles image generation before you install anything. It needs a browser with WebGPU support.

The tools

• ComfyUI builds image pipelines from connected nodes. It's the most flexible, and usually the first to support new models.
• Stable Diffusion WebUI (AUTOMATIC1111) and its forks, like Forge, give you a classic form-style interface.
• InvokeAI is a polished app with a canvas for editing.
• Draw Things is a popular native app on Mac.

All of these are free and run locally. ComfyUI or a Forge-style WebUI is a good first pick on a PC.

Getting good results

• Use the size a model was trained on. SDXL and Flux are built for around 1024×1024; SD 1.5 for 512×512. Far off that and you get stretched or duplicated subjects.
• Steps and guidance matter. Each model has recommended settings on its model card; start there. See how to read a model card.
• Fine-tunes and LoRAs add styles and subjects. They must match the base model family.
• Seeds let you repeat an image and tweak one thing at a time.

Licenses vary more than for text models

Image models come under a wide range of licenses, and some popular ones are for non-commercial use only. Fine-tunes inherit the base model's terms. Check before you use images commercially. See open model licenses explained.

Images and chat on one card

A 24 GB card can hold a chat model or an image model comfortably, but not both large ones at once. Ollama unloads idle models after a while, or you can stop one before starting the other. Here's what a 3090 runs on the chat side:

For a step-by-step setup, see the local image generation guide. Or open the Studio and make your first image in a minute.