Stable Diffusion 3.5 Medium VRAM Requirements
Stable Diffusion 3.5 Medium needs about 6.4 GB of VRAM with FP8 weights at 1024×1024, loading one model at a time. In FP16 it needs 11.6 GB (19.1 GB with everything loaded), and GGUF Q4 brings it to 6.4 GB. The smallest consumer GPU that runs it is the RTX 3050 8GB with Q4 weights, or the RTX 3050 8GB in FP8.
6.4 GB peak VRAM
Stable Diffusion 3.5 Medium at 1024×1024 with FP8 weights, loading one model at a time (ComfyUI default on smaller cards). Keeping everything loaded needs 11.6 GB.
- Diffusion model 2.3 GB
- Text encoders 5.2 GB
- VAE 0.2 GB
- Peak activations 3.1 GB
- Overhead 0.8 GB
GPU fit
- RTX 4060 8GBFitswith swapping
- RTX 3060 12GBFitswith swapping
- RTX 4060 Ti 16GBFitsall loaded
- RTX 4090 24GBFitsall loaded
- RTX 5090 32GBFitsall loaded
- RTX 6000 Ada 48GBFitsall loaded
- RTX PRO 6000 Blackwell 96GBFitsall loaded
- Apple M4 Max 128GBFitsall loaded
The VAE decode is the peak here. Turn on tiled VAE decoding to cut it to about 0.8 GB.
Stable Diffusion 3.5 Medium VRAM by precision
At 1024×1024. Text encoders in FP8 except for the FP16 row.
| Weights | Diffusion model | Peak (swapping) | All loaded | Smallest tier |
|---|---|---|---|---|
| FP16 / BF16 | 4.7 GB | 11.6 GB | 19.1 GB | 16 GB |
| FP8 | 2.3 GB | 6.4 GB | 11.6 GB | 8 GB |
| GGUF Q8_0 | 2.5 GB | 6.4 GB | 11.8 GB | 8 GB |
| GGUF Q5_K_S | 1.6 GB | 6.4 GB | 10.9 GB | 8 GB |
| GGUF Q4_K_S / NF4 | 1.3 GB | 6.4 GB | 10.6 GB | 8 GB |
VRAM by resolution
FP8 weights, with model swapping.
| Resolution | Latent tokens | Activations | Peak VRAM |
|---|---|---|---|
| 1024×1024 | 4,096 | 0.2 GB | 6.4 GB |
| 1536×1536 | 9,216 | 0.4 GB | 8.0 GB |
About Stable Diffusion 3.5 Medium
A 2.5B diffusion transformer that is small on its own, but its T5-XXL text encoder is almost twice its size. Loading T5 in FP8, or dropping it entirely, is how SD 3.5 Medium fits on 8 GB cards.
| Component | Size |
|---|---|
| Diffusion model | 2.5B parameters (MMDiT-X) |
| Text encoders | CLIP ViT-L (0.123B), OpenCLIP ViT-bigG (0.695B), T5-XXL encoder (4.76B) |
| VAE | 0.084B parameters |
| Classifier-free guidance | Yes (2 passes per step) |
| Released | 2024-10 |
| License | Stability AI Community License |
Stable Diffusion 3.5 Medium VRAM: common questions
Can SD 3.5 Medium run without the T5 encoder?
Yes. Running only the CLIP encoders saves about 9 GB in FP16 at some cost to prompt following.
How much VRAM does Stable Diffusion 3.5 Medium need?
At its default 1024×1024: about 11.6 GB in FP16 with model swapping (19.1 GB with everything loaded), 6.4 GB in FP8 and 6.4 GB with GGUF Q4 weights.