You don't need an expensive card for image generation. A 24GB GPU runs Flux with ControlNet for $0.12–0.34/hr. Here's the GPU each workflow actually needs, dated.
This post contains affiliate links. If you buy through them, Aliteq may earn a commission — at no extra cost to you. Prices verified at publish time.
Share
Here's the good news if you want to run ComfyUI, Stable Diffusion or Flux in the cloud: it's cheap, and you do not need an expensive card. A 24GB GPU handles Flux with ControlNet and LoRAs comfortably, and as of 22 Sep 2026 you can rent one — an RTX 3090 or 4090 — for $0.12 to $0.34 an hour. People routinely over-rent H100s for image work that a $0.14/hour card does just as well. Here's the GPU you actually need for each workflow, and the cheapest cloud way to run it. This is a spoke of our cloud-GPU pricing pillar.
How much VRAM you actually need
ComfyUI itself is light — the model you load decides the VRAM. SDXL sits around 8–12GB. Flux.1-dev in fp16 wants 16–24GB, and once you stack ControlNet and a couple of LoRAs you're using the full 24GB, which is why the RTX 3090/4090 (both 24GB) is the practical minimum for no-compromise Flux work; FP8 quantisation squeezes it onto a 16GB card if you're patient. Video models and long multi-stage pipelines are the only image-side workloads that justify a 48GB card. Match the card to the workflow and the bill stays tiny.
Cheapest cloud GPU for image gen · $/hr · captured 22 Sep 2026
SDXL (base + refiner)
Workflow
8–12GB
VRAM
RTX 3090 24GB
Cheapest card
$0.12 / $0.22
Flux + ControlNet + LoRA
Workflow
24GB
VRAM
RTX 3090 / 4090
Cheapest card
$0.12–0.14 / $0.22–0.34
Flux, faster / headroom
Workflow
32GB
VRAM
RTX 5090
Cheapest card
$0.40 / $0.69
Video / heavy pipelines
Workflow
48GB
VRAM
L40S / A6000
Cheapest card
$0.52 / $0.79
Workflow
VRAM
Cheapest card
Vast spot / RunPod
SDXL (base + refiner)
8–12GB
RTX 3090 24GB
$0.12 / $0.22
Flux + ControlNet + LoRA
24GB
RTX 3090 / 4090
$0.12–0.14 / $0.22–0.34
Flux, faster / headroom
32GB
RTX 5090
$0.40 / $0.69
Video / heavy pipelines
48GB
L40S / A6000
$0.52 / $0.79
Referral link
Run ComfyUI + Flux on a 24GB card from ~$0.14/hr
The cheapest way to self-host image generation: an RTX 3090 or 4090 on Vast.ai spot as of 22 Sep 2026. Plenty for Flux with ControlNet — no need to pay H100 rates for image work.
Referral link — we may earn a commission at no cost to you. Prices on our compare page are the provider's live figures, cheapest first; this never changes the ranking.
The cheap way to set it up
The practical workflow: spin up a 24GB instance from a ComfyUI template (both Vast.ai and RunPod have community images with ComfyUI and the common models preloaded), generate for an hour or two, then shut it down. At ~$0.14–0.34/hour, a solid image session costs pennies to a couple of dollars, and you only pay while it's running. Use a persistent network volume for your models and outputs so you're not re-downloading multi-gigabyte checkpoints on every spin-up — the small storage fee is far cheaper than the wasted GPU time of repeated cold downloads. On spot, checkpoint or export your outputs regularly in case the instance is reclaimed.
Prefer a persistent ComfyUI workspace? RunPod pods + network volumes.Referral link
An RTX 3090 (24GB) at around $0.12/hour on Vast.ai spot or $0.22/hour on RunPod (22 Sep 2026), or an RTX 4090 for a little more. Both have the 24GB you need for Flux with ControlNet and LoRAs. There's no reason to rent an H100 for image generation — the extra VRAM and throughput do nothing for SDXL or Flux that a 24GB card can't, so the cheaper card is the smarter choice.
How much VRAM do I need to run Flux in the cloud?
16–24GB. Flux.1-dev in fp16 needs about 16–24GB, and once you add ControlNet and LoRAs you'll want the full 24GB — so an RTX 3090 or 4090 is the practical minimum for no-compromise Flux. FP8 quantisation brings simpler Flux workflows onto a 16GB card at some quality/speed cost. SDXL is lighter, at 8–12GB.
How much does it cost to generate images on a cloud GPU?
Very little. At $0.12–0.34/hour for a 24GB card, an hour or two of ComfyUI work costs pennies to a couple of dollars, and you only pay while the instance runs. The main way people overspend is leaving the box on after they finish, or re-downloading large models on every spin-up — use a persistent volume and shut down promptly.
Should I rent a cloud GPU or use a hosted image tool?
Rent and self-host ComfyUI if you need custom nodes, private workflows, or the lowest per-image cost at volume. Use a hosted tool if you'd rather not manage a GPU at all and value speed of setup. For occasional, casual image generation a hosted service is simpler; for a repeatable production pipeline, a cheap rented 24GB card is hard to beat on cost.
Image generation is one of the cheapest things you can rent a GPU for — a 24GB card runs Flux and SDXL for pennies an hour, and an H100 is wasted money here. Compare live prices on our cloud-GPU compare page, read the pricing pillar, see the cheapest cloud GPU overall, and for hosted image tools the AI Money guides.
