Use case · diffusion

Image and video generation.

Diffusion models are small; the pipelines around them are not. What matters is a card with enough memory for the refiner, the ControlNets and the batch — and enough bandwidth to keep the steps moving.

Deploy ComfyUI See consumer GPUs

Entry card: $0.05/hRTX 3090, 24 GB, on spot
Fastest per $: RTX 50901,792 GB/s at $0.09/h
Template: ComfyUIInterface and model manager, pre-pulled
A 4-hour batch: $0.20On spot, billed per minute

What each pipeline needs

Weights are the easy part.

Parameter counts are public; the memory below is the weights at FP16 with our usual 20% headroom. Add the refiner, the upscaler, the ControlNets and the batch size — which is why 24 GB is the practical floor and 32 GB the comfortable one.

Pipeline	Parameters	Weights at FP16	Cheapest card that fits	Spot
SDXL 1.0Still the workhorse for 1024×1024 images	3.5B	7 GB	RTX A4000 · 16 GB	$0.03/h
FLUX.1 [dev]Large transformer model, much better prompt adherence	12B	24 GB	A40 · 48 GB	$0.06/h
SD 3.5 LargeText rendering and composition	8.1B	16 GB	A30 · 24 GB	$0.05/h
Video diffusion (5B class)Short clips; frames multiply the activation memory	5B	10 GB	RTX A4000 · 16 GB	$0.03/h

"Cheapest card that fits" is the weights only — the honest floor. For real pipelines take a 24 GB card at least, and a 32 GB one if you batch or chain models.

The cards people use

Consumer boards, rented by the minute.

Diffusion is where consumer hardware is unbeatable on price: no ECC, no NVLink, and the highest bandwidth per dollar in the catalogue.

All 47 models

GPU	Memory	Bandwidth	Spot	A 4-hour batch	Action
RTX 3090Ampere · 128 available	24 GB	936 GB/s	$0.05	$0.20	Deploy
RTX 4090Ada Lovelace · 128 available	24 GB	1,008 GB/s	$0.06	$0.24	Deploy
RTX 3090 TiAmpere · 24 available	24 GB	1,008 GB/s	$0.06	$0.24	Deploy
RTX 5090Blackwell · 128 available	32 GB	1,792 GB/s	$0.09	$0.36	Deploy

Need ECC memory, more than 32 GB or several models resident at once? The workstation cards — RTX 6000 Ada and RTX PRO 6000 — carry 48 GB and 96 GB.

Getting set up

Models on the volume, cache on scratch.

The pattern that avoids re-downloading twenty gigabytes of checkpoints every time an instance starts — and keeps your outputs when a spot instance is reclaimed.

inside the instance

# Models and outputs on the persistent volume.
$ ls /mnt/vol/models/checkpoints /mnt/vol/output

# Point ComfyUI at them; the interface URL appears in the console.
$ python main.py --listen 0.0.0.0 --port 8188 \
      --output-directory /mnt/vol/output \
      --extra-model-paths-config /mnt/vol/models/paths.yaml

# A batch, from the API, writing each image as it finishes.
$ curl -X POST http://localhost:8188/prompt -d @workflow.json

Which tier

A batch of images: Spot. Write each output to the volume as it finishes and a reclaim costs you one image.
An interactive session: On-demand. You are sitting in front of it; 2 minutes of warning does not help.
A public endpoint: On-demand, with the 99.9% monthly target. Put the overflow on spot behind a queue.
Model storage: A 250 GB volume holds a serious model library for $20.00 a month, and survives every instance.

See the ComfyUI template

FAQ

Generation questions

How much VRAM does image generation need?

Less than people expect for the weights — SDXL is about 3.5B parameters, roughly 7 GB at FP16 — and more than expected for everything else. Activations at high resolution, a refiner, ControlNets, upscalers and a batch all add up. 24 GB is comfortable for most work; 32 GB removes the need to think about it.

Which GPU gives the most images per dollar?

A consumer card, almost always. The RTX 5090 at $0.09 an hour on spot has 1,792 GB/s of bandwidth — more than most data-center parts — and diffusion is bandwidth-hungry. Data-center cards win when you need more than 32 GB or ECC memory.

Can I run ComfyUI in the browser?

Yes. The ComfyUI template starts the interface and the console shows its URL. Keep your models and outputs on a persistent volume so they survive the instance.

Is spot capacity sensible for image work?

For batches, absolutely: a reclaim costs you the image being generated, not the batch, as long as you write outputs to the volume as they finish. For an interactive session you are sitting in front of, take on-demand — being interrupted mid-prompt is annoying at any price.

Do you allow generated content of any kind?

Within the acceptable use policy. Illegal material — child sexual abuse material above all — ends the account and is reported. Beyond that, what you generate is your business, and we do not inspect it.

Get started

A 24 GB card and ComfyUI for $0.05 an hour.

Pay as you go — no contracts, no minimum commitment. Add credit, launch, stop whenever you want.

Deploy ComfyUI Other use cases

Billed per minute from the moment the instance is reachable. Minimum credit $40, no subscription.