Home/Solutions/Image and video

Start, generate, stop. That is all you pay for.

Stable Diffusion, ComfyUI graphs, video diffusion and classic render: the work is short and intense by nature. This is exactly where per-second billing pays off most.

You pay per second and stop whenever you want. No commitment.

Your work looks like this

Twenty minutes of generation, three hours of idle card.

Image and video work does not draw a flat load curve: an idea arrives, the card boils for a while, then you look at the results. The card does not need to keep running while you look.

Short bursts, long gaps

A batch takes 40 seconds, a video clip 12 minutes. A meter that keeps running through the review breaks in between costs more, by the end of the month, than the work itself.

Downloading models on every start

Checkpoints, LoRAs, VAEs and ControlNet weights run to tens of gigabytes. Pulling them again on every pod burns both time and traffic.

The forgotten pod

A ComfyUI session left open at night stays open until morning. Instead of leaving that to good intentions, we tie it to a cap: when the limit is reached the job pauses.

Why the second matters

The same work, two different ways of billing it.

The calculation below is made up and illustrative: eight separate generation sessions in one day, each averaging 12 minutes. Total real usage is 1 hour 36 minutes. The left column is what we measure; the right one answers the question of what would happen in a hypothetical model that rounds up to the hour.

Example day (made up)Per-second billingIf hours were rounded up (hypothetical)
8 sessions × 12 minutes96 minutes8 × 1 hour = 480 minutes
1× RTX 4090 · $0.42/hr (example)$0.67$3.36
20 working days a month$13.44$67.20

An example and illustrative calculation; it points at no one. The right-hand column is not a competitor price but a hypothetical billing model that rounds up to the hour. Our measurement is per second: if it ran for 12 minutes 40 seconds, you pay for 12 minutes 40 seconds.

The flow

One generation session from start to finish

Set up the model store once

Put checkpoints, LoRAs and ControlNet weights on a persistent volume. The volume lives independently of the pod; on the next start there is no download, only an attach.

Pick the card to match the job

24 GB is usually enough for SDXL at 1024 px; for video diffusion, long queues and ComfyUI graphs that want 48 GB you step up a tier.

Start the pod, connect its interface

ComfyUI, A1111 or your own image. You expose the port through the pod connection; Jupyter and the terminal stay on the same screen.

Generate, write the output to the volume

Results stay on the volume. When you stop the pod your output does not go with it; you download it with a signed link whenever you want.

Stop when you are done

The compute charge stops at shutdown. The only line that keeps running is storage, and it appears as a separate row on the invoice — we do not hide that row.

Set the cap so you do not have to remember

Whatever your daily working habits are, the hard cap stands behind you: when it fills, the pod is paused, not deleted.

Made-up scenario

A typical render day.

Not a real customer: an illustrative example constructed to show the product. The numbers are examples.

The work: a three-person design team producing campaign visuals with SDXL + ControlNet.
Storage: a 300 GB persistent volume; the checkpoint and LoRA archive is uploaded once.
The card: 1× RTX 4090 during the day (example $0.42/hr), an L40S when a video test is needed (example $0.79/hr).
The cap: a $150 monthly hard cap for the team; a notice at 80%, a pause at 100%.
CLIPythoncurl
# The model store once, then attached every time $ kaldera volume create sd-models --size 300 # Start the ComfyUI pod $ kaldera pod create --gpu RTX4090 \ --image comfyui/comfyui --volume sd-models:300 \ --port 8188 ✓ pod ready (1m 42s) → interface /p/8188 # Job done: stop it, keep the storage $ kaldera pod stop last compute stopped 1h 36m example $0.67 storage running 300 GB separate line

Choosing a card

VRAM and relative speed.

In image work VRAM decides your resolution and how complex a graph you can run; speed decides the rhythm of iteration. The speed column below is a relative classification — we do not publish numeric benchmarks, because our scorecard bot has only just started to accumulate measurements. Rather than inventing numbers, we prefer to give an ordering.

GPUVRAMComfortable workload (example)Relative speed classExample price
RTX 309024 GBSD 1.5 and SDXL at 1024 px, single user, experimentsBaseline$0.21/hr
RTX 409024 GBSDXL + ControlNet, fast iteration, batch generationHigh$0.42/hr
L40S48 GBVideo diffusion, ComfyUI graphs that want 48 GBHigh · large VRAM$0.79/hr
RTX A600048 GBLarge scene renders, high-resolution outputMedium · large VRAM$0.88/hr
H100 SXM80 GBHeavy video models, bulk generation in a short windowHighest$2.19/hr

Example prices, illustrative only; they move with supply and demand on the marketplace. "Relative speed class" is not a measurement result but a rough sense of direction. For storage and traffic line items see Pricing.

Model store

Keeping your checkpoint archive independent of the pod: Persistent Storage.

Pods and the marketplace

Picking offers by card, region and scorecard: GPU Pods.

Honest limits

What exists today, and what does not.

ComfyUI, A1111 or Blender pods with your own imagePhase 0 — available today
Persistent model store and output folderPhase 0 — available today
Per-second billing, hard cap, live meterPhase 0 — available today
One-click ComfyUI templates (model catalogue)Soon — no date fixed
Production endpoint that scales to zeroPhase 2 — soon
Batch render job queueSoon — no date fixed

Today you bring your own image instead of a ready-made template; that is a gap and we write it as one. The model catalogue sits in the nav with a "soon" badge — we do not describe it as if it already existed.

Frequently asked

Will I have to download my models again on every start?

No. A persistent volume is independent of the pod: you upload the checkpoint and LoRA archive once, then attach the same volume to later pods. Even if the pod is deleted the volume stays; only the storage line keeps running, and it appears as a separate row on the invoice.

How long does a pod take to start?

Our target is under two minutes. The actual time depends on the size of the image and the host network; the console records the real duration of every start in the event log. We do not yet have enough measurements to publish an average "start time" figure.

How fast is it? Can you give seconds per image?

Not today. Our benchmark bot tests every host regularly and the scorecard is fed by those measurements; but the data is not yet enough for a publishable, reproducible images-per-second table. Rather than inventing numbers, we give a relative classification.

Commercial use and content liability?

The content you generate is yours; the licence of the model and checkpoint you use is your responsibility — we do not provide model licences. Platform rules are written in the terms: Terms of Use.

What happens if I forget a pod?

The hard budget cap steps in: you get a notice at 80% and 95%, and at 100% the pod is paused with its disk preserved. The cap is a brake, not a guillotine — raise the limit and carry on where you left off.

Get started

Hold the card only as long as the work lasts. The rest is our problem.

You don't pay for seconds that don't work, a hard budget cap, per-second billing. Set up your model store once, then just generate.

The console runs on the local development address: localhost:3000