How to Run FLUX and Stable Diffusion XL on a Rented Cloud GPU (2026 Beginner Guide)

October 19, 2026 · 6 min read · By the DollarWise team

On this page

You don’t need a $2,000 graphics card to run state-of-the-art image models. For the price of a coffee, you can rent a cloud GPU by the hour and generate images with FLUX.1 or Stable Diffusion XL from your laptop. This beginner guide walks you through choosing the right GPU, understanding pricing, and getting your first images generated in about 30 minutes.

Disclosure: links to RunPod in this post are referral links — we may earn credits at no cost to you.

Why rent instead of buy?

A local RTX 4090 costs $1,600+ and ties up your machine for hours. Renting flips the math: you pay only for the hours you actually generate. A weekend of heavy experimenting might cost $5–$15 total — less than a month of most hosted AI image subscriptions, with the freedom to run any open model, any custom checkpoint, and your own workflows.

Step 1: pick your model and its VRAM requirement

VRAM is the hard constraint. The model must fit in the GPU’s memory, full stop. Here are the numbers that matter in 2026:

FLUX.1 dev (fp16): ~24GB VRAM. Needs an RTX 4090 (24GB), RTX 3090 (24GB), or A100 (40/80GB). The fp8-quantized version squeezes onto 16GB cards like the RTX 4080 with slightly reduced quality.
FLUX.1 schnell: the distilled, faster variant — same VRAM needs, but 4 steps instead of 20–50, so it generates much faster per image.
Stable Diffusion XL (SDXL 1.0): ~12GB VRAM. Comfortable on an RTX 3060 (12GB), RTX 4070 (12GB), RTX 3090, or A4000. The most forgiving option for beginners and the cheapest to run.
SDXL Turbo / Lightning: distilled SDXL variants that generate in 1–8 steps — great on 12GB cards for fast iteration.

Rule of thumb: if you’re new, start with SDXL on a 12–16GB card. Move to FLUX on a 24GB card when you want its noticeably better prompt adherence and text rendering.

Step 2: understand spot vs on-demand pricing

Cloud GPU platforms like RunPod offer two pricing models:

Community/spot GPUs: spare capacity from individual providers. An RTX 3090 typically runs $0.20–$0.35/hr, an RTX 4090 around $0.40–$0.80/hr. The catch: they can be interrupted if the provider reclaims the machine. Fine for experimentation; risky for 6-hour training runs.
Secure/on-demand GPUs: datacenter hardware with guaranteed availability. Expect roughly 2–3x spot prices — an A100 40GB around $1–$2/hr, an H100 higher. Use these when uptime matters.

Storage costs a few cents per GB per month, and most platforms charge a small fee for keeping a stopped pod’s disk around. Delete pods and volumes you’re not using.

Step 3: launch your first pod (RunPod walkthrough)

Here’s the standard flow, which is similar on most platforms:

Create an account and add funds. Add at least $10 to unlock the full GPU marketplace and earn referral credits (more on those below).
Deploy a pod. Choose your GPU (RTX 3090 for SDXL, RTX 4090 for FLUX), pick a PyTorch or Stable Diffusion template, and select enough container disk (30–50GB for model weights).
Connect. Open the pod’s web terminal or the bundled JupyterLab / ComfyUI / Automatic1111 interface, depending on the template.
Download the model. Pull weights from Hugging Face (huggingface-cli download or a direct wget). FLUX.1 dev is ~23GB; SDXL base is ~7GB. First download takes a while — subsequent launches reuse the volume.
Generate. In ComfyUI or Automatic1111, load the checkpoint, type a prompt, and hit generate. Your first image should appear within a minute of the model loading.
Stop the pod when done. This is the step beginners forget. Billing is per-second while the pod runs — always stop it when you walk away.

Real cost examples

What does a typical session actually cost? Assuming an RTX 3090 at $0.30/hr (community cloud):

100 SDXL images (a few hours of casual prompting and tweaking): roughly $1–$2 of GPU time.
100 FLUX images on an RTX 4090 at ~$0.70/hr: roughly $2–$4, since FLUX needs more steps per image.
A full weekend project (say 10 active hours on a 3090): about $3.

Compare that to $20–$30/month for hosted generators, and renting wins whenever you generate in bursts rather than daily.

About RunPod’s referral credits (the honest version)

Since we’re linking RunPod above, here’s exactly how their referral program works so you can decide if it’s worth it:

Rewards are credits, not cash. Both the referrer and the referred person get a randomized $5–$500 credit bonus once the referred account spends its first $10.
The referrer also earns 3% of the referral’s Pod spend and 5% of Serverless spend for the referral’s first 6 months — again, as credits.
Credits are non-transferable and expire after 90 days.
Actual cash payouts exist only at RunPod’s invited affiliate tier (25 paying referrals) — not something a normal signup reaches.

So the honest reader incentive is: $5–$500 in free GPU credits for both of us once you add your first $10. Nothing more, nothing promised beyond that. If you’d rather skip the referral entirely, just visit runpod.io directly — the pricing and product are identical either way.

Troubleshooting tips for beginners

Out of memory errors: drop to a quantized model (fp8), reduce image resolution, or step up to a bigger GPU.
Slow first generation: the model loads into VRAM on first use — this is normal, later images are fast.
Pod got interrupted (spot): relaunch and reload; keep your prompts and workflows saved externally (GitHub, or download the volume).
Costs creeping up: check for stopped-but-not-deleted pods holding paid storage, and set billing alerts.

Frequently asked questions

How much VRAM do I need to run FLUX.1?

FLUX.1 dev needs about 24GB of VRAM in fp16, so an RTX 4090 or RTX 3090 works. Quantized versions (fp8) can run on 16GB cards like the RTX 4080.

Can I run Stable Diffusion XL on a cheap cloud GPU?

Yes. SDXL runs comfortably on ~12GB of VRAM, so cards like the RTX 3060, RTX 4070, or A4000 handle it, often under $0.20/hr on community cloud platforms.

What is the difference between spot and on-demand GPU pricing?

On-demand (secure cloud) GPUs stay available until you stop them but cost more. Spot (community cloud) GPUs are cheaper — often 50-70% less — but can be interrupted when demand spikes.

How much does it cost to generate 100 images on a rented GPU?

On a $0.30/hr RTX 3090, 100 SDXL images (a few hours of casual use) typically costs $1-2 total. FLUX on a $0.80/hr RTX 4090 runs roughly $2-4 for the same session.

#cloud gpu #flux #stable diffusion #ai image generation

Written by the DollarWise team

We turn confusing money topics into plain-English guides for beginners. Every article is written to be actionable on day one — no jargon, no hype, no sales pitch.