GPU for Stable Diffusion: how much VRAM each model needs, and what it costs in crypto
Image models have grown from 1 billion parameters to 12 billion in three years, and memory now decides which card works; each card here is rented with crypto only. Here is what each model needs, which card to rent for it, and how to work out the cost per image.
The GPU you need for Stable Diffusion depends on the model. SD 1.5 and SDXL run comfortably on any NVIDIA card with 12 GB or more, and a 24 GB RTX 4090 handles them with large batches and LoRA training. Flux.1 and SD 3.5 Large need 24 GB with FP8 weights, or 32 GB to run Flux in BF16, which points to an RTX 5090. For video models and heavy ComfyUI pipelines, 48 GB on an L40S removes most limits. Renting, an RTX 4090 costs $0.34 an hour here, which at 5 seconds per image is about $0.47 per 1,000 images.
How much GPU memory each image model needs
Model sizes from each model's official card. Weights take about 2 bytes per parameter in FP16 or BF16 and 1 byte in FP8; text encoders, the VAE and the image being generated need memory on top.
| Model | Official size | Memory for the weights | Card to rent |
|---|---|---|---|
| Stable Diffusion 1.5 | 860M-parameter U-Net, 123M text encoder | About 2 GB in FP16 | Any card; RTX 4090 for big batches |
| SDXL 1.0 | 3.5B-parameter base model | About 7 GB in FP16 | RTX 4090 |
| SD 3.5 Medium | 2.5B parameters | About 5 GB, plus three text encoders | RTX 4090 |
| SD 3.5 Large | 8B parameters | About 16 GB in BF16, plus text encoders (T5-XXL alone is about 9.5 GB in FP16) | RTX 5090, or RTX 4090 with FP8 or quantized encoders |
| Flux.1 dev or schnell | 12B parameters | About 24 GB in BF16, 12 GB in FP8, plus T5-XXL | RTX 4090 in FP8; RTX 5090 in BF16 with an FP8 encoder |
| Wan 2.1 text-to-video 1.3B | 1.3B parameters | 8.19 GB of VRAM, according to the Wan team | Any card we rent |
| Wan 2.1 14B | 14B parameters | About 28 GB in BF16, 14 GB in FP8, plus encoders | RTX 5090 in FP8, L40S or A100 |
Higher resolutions, larger batches, ControlNets and upscalers all add memory. When a job runs out, ComfyUI and Diffusers can offload parts of the model to system memory, at a large cost in speed.
Which card to rent for each job
- RTX 4090, 24 GB, $0.34 an hour. The best GPU for Stable Diffusion on a budget: SD 1.5 and SDXL with batches, SDXL LoRA training, SD 3.5 Medium, and Flux.1 with FP8 weights. Most single-image ComfyUI workflows fit.
- RTX 5090, 32 GB, $0.49 an hour. The extra 8 GB lets Flux.1 dev run in BF16 next to an FP8 text encoder, and SD 3.5 Large run without heavy quantization. Its memory bandwidth, 1,792 GB/s against 1,008 GB/s, also speeds up every step.
- L40S, 48 GB, $0.87 an hour. For video models such as Wan 2.1 14B, long ComfyUI graphs that keep several models loaded, and Flux LoRA training without squeezing.
- A100 and H100. Rarely worth it for image generation alone. They make sense for training larger models or serving many users at once from one machine.
All of them are whole cards, 1 to 8 per machine, billed by the minute. For a ready-to-use setup, see ComfyUI on a rented GPU.
Cost per 1,000 images
What 1,000 images cost at our list prices, for the seconds each image takes on your workflow. Measure your own time per image, then read across.
| Card, per hour | 2 s per image | 5 s per image | 10 s per image | 30 s per image |
|---|---|---|---|---|
| RTX 4090, $0.34 | $0.19 | $0.47 | $0.94 | $2.83 |
| RTX 5090, $0.49 | $0.27 | $0.68 | $1.36 | $4.08 |
| L40S, $0.87 | $0.48 | $1.21 | $2.42 | $7.25 |
Cost = hourly price ÷ 3,600 × seconds per image × 1,000. A faster card is cheaper per image only if it saves time in proportion to its price: the RTX 5090 must be 1.44 times faster than the RTX 4090 to cost less per image. Prices on what a GPU-hour costs.
Image work usually runs on an RTX 4090 or an RTX 5090: 75 hours for $25.50 or 60 hours for $29.40, paid only in crypto, from USDT to Monero. No card is kept on file and the hours never expire. See how to pay with USDT or Bitcoin.
Measure your own seconds per image
Published benchmarks rarely match your model, resolution, sampler and step count. Ten minutes on a rented card gives you a real number. ComfyUI prints the time of each run in its console; with Diffusers, time a loop like this one:
import time, torch
from diffusers import StableDiffusionXLPipeline
pipe = StableDiffusionXLPipeline.from_pretrained(
"stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
).to("cuda")
pipe("warm-up", num_inference_steps=5) # first run loads kernels
start = time.time()
for _ in range(20):
pipe("a lighthouse at dawn, 35mm photo", num_inference_steps=30)
print(f"{(time.time() - start) / 20:.2f} s per image")
print(f"{torch.cuda.max_memory_allocated() / 1e9:.1f} GB peak VRAM")Training LoRAs
Training needs more memory than generating, because gradients and optimizer states sit next to the model. SDXL LoRA training typically fits in 12 to 16 GB, so an RTX 4090 trains with room for larger batches. Flux LoRA training tools generally ask for 24 GB or more with the base model quantized, which makes the RTX 5090 or the L40S the comfortable choice. Training runs are where per-minute billing helps: a LoRA that trains in two hours costs two hours, and the machine can be stopped as soon as the last checkpoint is saved.
Licences to check
Renting a GPU does not change the licence of the model you run on it.
- Flux.1 dev is released under a non-commercial licence for the model, while its page says generated images can be used for personal, scientific and commercial purposes. Flux.1 schnell uses a more permissive licence; read the one on its own model card.
- SD 3.5 models use Stability AI's Community License: free for research, non-commercial and commercial use for individuals and organizations under $1 million in annual revenue, with an Enterprise licence above that.
- SDXL and SD 1.5 use their own open licences with use restrictions; check them before building a product.
Running a model as a paid service for others usually needs more than generating images for yourself, so read the terms that apply to your case.
Buy a card or rent one
If you generate images every day for years, a card of your own pays for itself. If you work in bursts, want to try Flux or video models before buying 32 GB or more, or need several cards for a deadline, renting costs less. A 24 GB RTX 4090 bought for $2,000 equals about 5,880 hours of GPU rental here at $0.34, before the cost of the computer around it and the power it draws.
Stable Diffusion GPU questions
What kind of computer do I need to run Stable Diffusion?
An NVIDIA GPU with at least 8 GB of VRAM runs SD 1.5; 12 GB or more is comfortable for SDXL, and 24 GB or more for Flux and SD 3.5 Large. Without such a card, a rented cloud GPU gives you the same result in a browser or over SSH.
Are GPUs needed for inference?
Not strictly: Stable Diffusion can run on a CPU, but each image takes far longer. A GPU is what makes generation fast enough to iterate on prompts.
Can I run Stable Diffusion on a cloud GPU?
Yes. Rent a machine with an RTX 4090 or RTX 5090, install ComfyUI or Diffusers on our PyTorch image, and open the interface through an SSH tunnel. Nothing runs on your own computer except the browser.
Do AMD cards work for Stable Diffusion?
Some do, through ROCm, but most tools, guides and custom nodes assume NVIDIA and CUDA. Every card we rent is NVIDIA.
Sources
- Stable Diffusion v1 model card (CompVis)read 19 November 2026
- Stability AI, SDXL 1.0 announcementread 19 November 2026
- Stable Diffusion 3.5 Large model cardread 19 November 2026
- Stable Diffusion 3.5 Medium model cardread 19 November 2026
- FLUX.1 [dev] model card (Black Forest Labs)read 19 November 2026
- Wan 2.1 repositoryread 19 November 2026