Skip to content
Local AI Wizard
Your machine
Theme

How much memory do I need to generate images locally?

What publishers document, and what fits a 16 GB graphics card.

Memory · updated

Short answer

It depends on the model. 8 of the 21 reviewed image models with a documented figure fit a 16 GB graphics card.

  • Example: FLUX.2 [klein] 4B documents a 13 GB path.
  • No documented figure: Plan about 2.4 GB per billion parameters, plus 2 GB.

Publisher-documented paths from each model page; the estimate is the site's 16-bit media formula. Formulas

Why images need more

A text model is one file you can shrink to 4 bits. An image model is usually several parts that run together: a text encoder that reads the prompt, a large transformer that builds the picture step by step, and a decoder (VAE) that turns the result into pixels. Publishers ship these in 16-bit, and the text encoder alone can be billions of parameters. Higher resolutions also need more working memory during generation.

So the site does not use the text model's Q4 rule for images. When a publisher documents a memory path (a GPU size on a platform, sometimes with offload or quantization), the page uses that figure. A “conditional” path is one that only works with such a trick, and it is slower.

Documented paths

The smallest path each publisher documents, and whether it fits RTX 4060 Ti · 16 GB VRAM. A CUDA figure is not carried over to a Mac.

Reviewed image models with a documented memory path: smallest path, platform, mode, fit on a 16 GB card and commercial use.
Model · parametersSmallest path (GB)Platform · mode16 GB cardCommercial use
OmniGen V2
VectorSpaceLab · 7B
3CUDA · conditionalFits, conditionalAllowed
Qwen-Image
Qwen · 27.1B
4CUDA · conditionalFits, conditionalAllowed
Stable Diffusion 3.5 Medium
Stability AI · 2.5B
9.9any · conditionalFits, conditionalAllowed with conditions
Stable Diffusion 3.5 Large
Stability AI · 8B
11CUDA · conditionalFits, conditionalAllowed with conditions
Bagel
ByteDance · 14B
12CUDA · conditionalFits, conditionalAllowed
FLUX.2 [klein] 4B
Black Forest Labs · 4B
13CUDA · directFitsAllowed
FLUX.2 [klein] Base 4B
Black Forest Labs · 4B
13CUDA · directFitsAllowed
Z-Image Turbo
Alibaba · 6B
16CUDA · directFitsAllowed
LongCat Image
Meituan · 6B
17CUDA · conditionalNeeds more memoryAllowed
GLM-Image
Z AI · 16B
23CUDA · conditionalNeeds more memoryAllowed
ERNIE Image
Baidu · 8B
24CUDA · directNeeds more memoryAllowed
ERNIE Image Turbo
Baidu · 8B
24CUDA · directNeeds more memoryAllowed
FLUX.2 [dev]
Black Forest Labs · 32B
24CUDA · conditionalNeeds more memoryNot allowed
HunyuanImage-2.1
Tencent Hunyuan · 17B
24CUDA · conditionalNeeds more memoryAllowed with conditions

Without a figure

18 of the 39 reviewed image models have no documented memory path. For them the site shows a labelled 16-bit estimate: 2 bytes per parameter, 20% working room and 2 GB reserve. FLUX.1 [dev], with 12 billion parameters, comes to about 31 GB that way. Community files in 8 or 4 bits and offload tools can bring that down, but each one is its own setup, not the publisher's figure.

Side by side: FLUX.2 klein 4B vs Z-Image Turbo for fast images · Stable Diffusion 3.5 Large vs FLUX.1 dev for images