What hardware do you need to generate images locally? File sizes of real open image models, and what each memory size can hold.
Short answer. You need a GPU with enough memory to hold the image model: about 7 GB for SDXL at 16-bit, about 34 GB for the full set of FLUX.1 [schnell] files and about 58 GB for Qwen-Image, before the working memory used while drawing. A 24 GB card runs SDXL comfortably and FLUX.1 schnell only with offloading or a smaller format; a 64 to 128 GB unified-memory machine can hold the larger models whole. Speed depends on the GPU’s compute and varies widely, so this page gives sizes, not timings. ANVE generates images on the box’s own GPU through ComfyUI.
01
File sizes of three open image models
Sizes are the file sizes in each model’s Hugging Face repository, in decimal gigabytes. Parameter counts come from the repositories’ metadata. The weights are not the whole story: the working memory used while an image is drawn grows with the image size, and this page does not quantify it.
Model
Main network
Text encoders
VAE
All files at 16-bit
Stable Diffusion XL base 1.0
5.1 GB (UNet, 2.6 billion parameters)
1.6 GB
0.3 GB
7.1 GB
FLUX.1 [schnell]
23.8 GB (transformer, 11.9 billion parameters)
9.8 GB (T5 and CLIP)
0.3 GB
33.9 GB
Qwen-Image
40.9 GB (transformer, 20.4 billion parameters)
16.6 GB
0.3 GB
57.7 GB
02
What each memory size can hold
AMD’s own notes agree on the direction: its FAQ says Stable Diffusion 3 Medium needs 18 GB of VRAM at 16-bit, that FLUX schnell can need 32 GB even in a partial-offload mode, and that a Ryzen AI Max+ 395 with 128 GB can run FLUX schnell with full offload. Halving the precision halves the weights, which is the same arithmetic as for language models.
Memory available
SDXL
FLUX.1 [schnell]
Qwen-Image
8 GB graphics card
Tight: the files alone are 7.1 GB
No
No
12 to 16 GB graphics card
Yes
Only with a smaller format or offloading to system memory
No
24 GB graphics card
Yes
The transformer alone is 23.8 GB, so offload or use 8-bit
No without a smaller format
96 GB of graphics memory (128 GB unified memory)
Yes
Yes, whole
Yes, whole, with room to spare
03
Software and operating systems
ComfyUI, a common local front end for image models, runs on Windows, Linux and macOS (Apple silicon). Its documentation lists NVIDIA GPUs, AMD GPUs (ROCm 7.2 on Linux, and Windows through AMD’s PyTorch packages), Intel Arc and Apple silicon, and a CPU mode that it labels slower. Check that your exact GPU is on the list before buying.
04
Speed, sharing and ANVE
Each image needs many passes through a large network, so the GPU’s arithmetic speed matters more here than for chat. Timings vary with model, image size and number of steps, so no figure is quoted. Run image jobs when chat is quiet, or give them their own memory.
ANVE generates images on the box’s own GPU through ComfyUI in its development build, with vision models for reading images separately. ANVE has not published image timings or the list of image models for its launch hardware.
05
Questions people ask
Can images be generated on a CPU?
ComfyUI has a CPU mode, started with a command-line parameter, and its documentation labels it slower. A GPU is the practical route for regular use.
Do I need an NVIDIA card?
No. ComfyUI’s documentation lists NVIDIA, AMD, Intel Arc and Apple silicon. NVIDIA is the most widely supported, but check the list for your exact card.
Is 24 GB enough for image generation?
For SDXL, comfortably: its files total about 7.1 GB. For FLUX.1 schnell it is tight, because the main network alone is 23.8 GB at 16-bit, so you would offload part of it or use a smaller format.
06
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
