How much VRAM does Stable Diffusion 3.5 Medium need?

At BF16 with everything resident, Stable Diffusion 3.5 Medium peaks at about 18.57 GB. Offloading components one at a time brings that to 5.84 GB, and quantising the weights to 4-bit brings it to 7.7 GB.

What is the cheapest GPU that runs Stable Diffusion 3.5 Medium?

The GeForce GTX 1650 at around $100, running it at BF16 with sequential offload, taking about 51 min 50 s per generation.

Stable Diffusion · Diffusion transformer

Stable Diffusion 3.5 Medium

Stable Diffusion 3.5 Medium is a pipeline of 5 networks totalling 8.135 billion parameters, of which 2.47 billion do the actual generating. At 1024x1024 that is 4,096 latent tokens per denoising step.

Whole pipeline8.135B

Denoiser2.47B

Download15.1 GB

Default steps28

What the pipeline is made of

This is the part that separates a media model from a language model: these are separate networks, and whether they sit in memory together is your choice, not the model's.

Component	Parameters	Published as	On disk	What it does
transformer	2.47B	BF16	4.6 GB	Does the generating. Runs once per step, so it dominates both memory and time.
text_encoder	0.124B	BF16	0.23 GB	Turns your prompt into conditioning. Runs once, then can leave the GPU entirely.
text_encoder_2	0.695B	BF16	1.29 GB	Turns your prompt into conditioning. Runs once, then can leave the GPU entirely.
text_encoder_3	4.762B	BF16	8.87 GB	Turns your prompt into conditioning. Runs once, then can leave the GPU entirely.
vae	0.084B	BF16	0.16 GB	Converts between pixels and the compressed latent space. Small, but its decode pass is a memory spike.

These figures were derived from the published file sizes, because the repository is gated.

Memory by precision

Everything resident, at the model's native output size, on a card large enough that precision is the only constraint.

Precision	Peak VRAM	Quality	What it costs you
BF16	18.57 GB	100%	How the weights are published. No loss, and the largest footprint.
FP8	11.07 GB	97%	Halves the denoiser with a small, usually invisible cost. Needs Ada or newer.
GGUF Q8_0	11.54 GB	98%	Works on any card, unlike FP8. Slightly slower than native precision.
GGUF Q5_K_M	8.9 GB	95%	A middle step when Q8 will not fit.
GGUF Q4_K_M	8.1 GB	91%	The usual way a 12B image model gets onto an 8 GB card. Detail softens.
NF4	7.7 GB	89%	Aggressive 4-bit. Fast to load, noticeably looser on fine detail.

Memory by arrangement

The other lever: the same weights at the same precision, moved around differently. On this model the strongest lever is neither: run the text encoder on the processor. It is 69% of the pipeline, and once it is off the GPU the 30% denoiser is all that remains.

Arrangement	Peak VRAM	Time	How it works
Everything resident	18.57 GB	9.5 s	All components stay on the GPU. Fastest, and needs the most memory.
Text encoder on CPU	5.84 GB	10 s	The prompt is encoded once on the processor, so the encoder never touches the GPU at all. Standard practice on video models, whose encoders are often larger than their denoisers.
Component offload	9.96 GB	11 s	One component on the GPU at a time. The text encoder runs, then makes way for the denoiser. Costs a few seconds per generation.
Sequential offload	3.82 GB	56 s	Weights stream layer by layer from system RAM. Runs almost anything on almost anything, and is many times slower.

What a bigger output costs

Memory climbs with the pixel count, and the VAE decode pass climbs with it.

Output	Latent tokens	Peak VRAM	Time
512 x 512	1,024	16.68 GB	1.6 s
768 x 768	2,304	17.47 GB	4.3 s
1024 x 1024	4,096	18.57 GB	9.5 s
1536 x 1536	9,216	21.71 GB	34 s
2048 x 2048	16,384	26.12 GB	1 min 30 s

Measured at BF16 with everything resident on an RTX 4090, so the columns compare with each other rather than with your machine.

Which hardware runs Stable Diffusion 3.5 Medium

118 of 118 consumer devices run it in some arrangement. Each row shows the fastest arrangement that fits on that device.

Device	Memory	Verdict	Arrangement	Peak	Time
RTX PRO 6000 Blackwell NVIDIA	96 GB	Workable	BF16 everything resident	18.57 GB	6.3 s
GeForce RTX 5090 NVIDIA	32 GB	Workable	BF16 everything resident	18.57 GB	7.6 s
RTX 6000 Ada Generation NVIDIA	48 GB	Workable	BF16 everything resident	18.57 GB	8.6 s
GeForce RTX 4090 NVIDIA	24 GB	Workable	BF16 everything resident	18.57 GB	9.5 s
GeForce RTX 5090 Laptop NVIDIA	24 GB	Workable	BF16 everything resident	18.57 GB	12 s
NVIDIA DGX Spark 128GB NVIDIA	128 GB	Workable	BF16 everything resident	18.57 GB	13 s
GeForce RTX 5080 NVIDIA	16 GB	Workable	GGUF Q8_0 everything resident	11.54 GB	14 s
GeForce RTX 4080 SUPER NVIDIA	16 GB	Workable	GGUF Q8_0 everything resident	11.54 GB	15 s
GeForce RTX 4090 Laptop NVIDIA	16 GB	Workable	GGUF Q8_0 everything resident	11.54 GB	16 s
GeForce RTX 4080 NVIDIA	16 GB	Workable	GGUF Q8_0 everything resident	11.54 GB	16 s
GeForce RTX 5080 Laptop NVIDIA	16 GB	Workable	GGUF Q8_0 everything resident	11.54 GB	16 s
GeForce RTX 5070 Ti NVIDIA	16 GB	Workable	GGUF Q8_0 everything resident	11.54 GB	18 s
GeForce RTX 4070 Ti SUPER NVIDIA	16 GB	Workable	GGUF Q8_0 everything resident	11.54 GB	18 s
GeForce RTX 4070 Ti NVIDIA	12 GB	Workable	FP8 everything resident	11.07 GB	19 s
GeForce RTX 3090 Ti NVIDIA	24 GB	Workable	BF16 everything resident	18.57 GB	19 s
RTX A6000 NVIDIA	48 GB	Workable	BF16 everything resident	18.57 GB	20 s
GeForce RTX 4080 Laptop NVIDIA	12 GB	Slow	FP8 everything resident	11.07 GB	21 s
GeForce RTX 3090 NVIDIA	24 GB	Slow	BF16 everything resident	18.57 GB	22 s
GeForce RTX 4070 SUPER NVIDIA	12 GB	Slow	FP8 everything resident	11.07 GB	22 s
GeForce RTX 3080 Ti NVIDIA	12 GB	Slow	FP8 everything resident	11.07 GB	23 s
GeForce RTX 5070 NVIDIA	12 GB	Slow	FP8 everything resident	11.07 GB	25 s
GeForce RTX 3080 12GB NVIDIA	12 GB	Slow	FP8 everything resident	11.07 GB	26 s
GeForce RTX 3080 10GB NVIDIA	10 GB	Slow	GGUF Q5_K_M everything resident	8.9 GB	26 s
GeForce RTX 4070 NVIDIA	12 GB	Slow	FP8 everything resident	11.07 GB	27 s
RTX A5000 NVIDIA	24 GB	Slow	BF16 everything resident	18.57 GB	28 s
GeForce RTX 2080 Ti NVIDIA	11 GB	Slow	GGUF Q5_K_M everything resident	8.9 GB	29 s
GeForce RTX 5060 Ti 16GB NVIDIA	16 GB	Slow	GGUF Q8_0 everything resident	11.54 GB	32 s
GeForce RTX 5060 Ti 8GB NVIDIA	8 GB	Slow	BF16 text encoder on cpu	5.84 GB	34 s
GeForce RTX 4070 Laptop NVIDIA	8 GB	Slow	BF16 text encoder on cpu	5.84 GB	35 s
GeForce RTX 4060 Ti 16GB NVIDIA	16 GB	Slow	GGUF Q8_0 everything resident	11.54 GB	35 s
Jetson AGX Orin 32GB NVIDIA	32 GB	Slow	BF16 everything resident	18.57 GB	36 s
Jetson AGX Orin 64GB NVIDIA	64 GB	Slow	BF16 everything resident	18.57 GB	36 s
Radeon RX 7900 XTX AMD	24 GB	Slow	BF16 everything resident	18.57 GB	37 s
GeForce RTX 4060 Ti 8GB NVIDIA	8 GB	Slow	BF16 text encoder on cpu	5.84 GB	37 s
Radeon PRO W7900 AMD	48 GB	Slow	BF16 everything resident	18.57 GB	37 s
GeForce RTX 3070 Ti NVIDIA	8 GB	Slow	BF16 text encoder on cpu	5.84 GB	37 s
GeForce RTX 3070 NVIDIA	8 GB	Slow	BF16 text encoder on cpu	5.84 GB	40 s
RTX A4000 NVIDIA	16 GB	Slow	GGUF Q8_0 everything resident	11.54 GB	40 s
GeForce RTX 5060 NVIDIA	8 GB	Slow	BF16 text encoder on cpu	5.84 GB	42 s
Radeon RX 7900 XT AMD	20 GB	Slow	BF16 everything resident	18.57 GB	44 s
Radeon RX 9070 XT AMD	16 GB	Slow	GGUF Q8_0 everything resident	11.54 GB	46 s
Radeon RX 7900 GRE AMD	16 GB	Slow	GGUF Q8_0 everything resident	11.54 GB	49 s
GeForce RTX 4060 Laptop NVIDIA	8 GB	Slow	BF16 text encoder on cpu	5.84 GB	49 s
GeForce RTX 3060 Ti NVIDIA	8 GB	Slow	BF16 text encoder on cpu	5.84 GB	50 s
GeForce RTX 4060 NVIDIA	8 GB	Slow	BF16 text encoder on cpu	5.84 GB	53 s

Source: stabilityai/stable-diffusion-3.5-medium . Downloaded 0.1 million times in the last month. See how we calculate, or browse every image model.