Skip to main content
NexGPU

AI image & video

Models your card cannot run, for cents a session

FLUX wants 24GB. Video diffusion models routinely want 40GB or more. Those numbers lock most local cards out. Renting one for a night costs far less than buying a 4090 for work you do intermittently.

Sign up freeView pricing
RTX 4090 per hour
$0.540
RTX 5090 per hour
$0.723
VRAM for full FLUX
24GB+

01 — When it applies

Why creators should rent

VRAM requirements for image generation have multiplied in two years. SD 1.5 was comfortable in 8GB, SDXL wants 12GB to run smoothly, FLUX.1 at full precision needs 24GB, and video diffusion models keep pushing higher. Buy a card to catch one wave and the next model generation locks you out again.

The more practical problem is utilisation. Creative work is bursty — an all-night session, then nothing for days. The price of a 4090 workstation buys thousands of cloud hours. At $0.540 per hour, 100 SDXL images takes about fifteen minutes and costs less than breakfast.

And you are not limited to one card. Run three in parallel for a batch style test and destroy them all when it finishes; swap up to an A6000 or A100 when you need more VRAM. Local hardware cannot give you that.

02 — Workloads

What creators run here

From single-image refinement to batch output, stills to motion.

  • Batch generation and style exploration

    One prompt across twenty seeds, ten samplers and five LoRA weights. An overnight job locally; three parallel instances finish it in half an hour here, for less money.

  • FLUX and other high-VRAM models

    FLUX.1 dev at full precision needs 24GB. A local 12GB card can only run quantised builds with visible quality loss. Rent a 4090 or 5090 and run the original.

  • AI video generation

    AnimateDiff, SVD and newer video diffusion models demand far more VRAM and compute than image models — 24GB minimum, and A6000-class is where it stops being cramped. Per-session rental is the only economical route.

  • LoRA training and character consistency

    Training a character or style LoRA typically runs one to three hours, about a dollar on a 4090. Download the weights when it finishes and destroy the instance.

03 — Stack

Creative environments, preinstalled

Ready on boot. No Python environment to configure, no xformers to compile, no CUDA version to look up.

  • ComfyUI — workflow first

    Node-based workflows with the common custom nodes and samplers already installed, covering SDXL, FLUX, video models and the ControlNet family. The right pick when you need precise control over the generation graph.

    ComfyUI · FLUX · SDXL · ControlNet

  • AUTOMATIC1111 — fastest to familiar

    The classic WebUI with the most mature extension ecosystem. Right when you already think in A1111, or need an extension that only exists there.

    A1111 · Extensions · img2img

  • Training and post-processing chain

    Kohya_ss and sd-scripts for LoRA training, plus upscaling, frame interpolation and matting models — chain the whole pipeline on one machine.

    Kohya_ss · Upscaling · Interpolation

04 — Choosing a GPU

Which card for which model

Image generation is sensitive to both VRAM and compute — VRAM decides whether it runs, compute decides how fast.

Use caseVRAM neededCheapest availableNotes
SD 1.5 and smaller models8GBGTX 1080$0.094/hrEnough to start. Fine for practice, small batches and existing legacy workflows.
SDXL production work12GBRTX 3060$0.100/hrComfortable for the full SDXL pipeline including the refiner, and noticeably faster than squeezing into 8GB.
Full FLUX.1 dev / LoRA training24GBTesla V100$0.188/hrThe current mainstream tier for serious work. Consumer flagships land at 24GB and give the best value per dollar.
Video generation / heavy parallelism48GBQ RTX 8000$0.508/hrVideo diffusion and multi-stream work need more VRAM headroom; at 48GB you stop tuning batch size constantly.

"Cheapest available" is derived from live inventory — the lowest per-GPU rate among models that clear the VRAM bar — and moves with the market. Multi-GPU nodes rent whole.

05 — Getting started

From order to first image

  • 01

    Pick a 24GB card

    RTX 4090 or 5090 offers the best value today. Size the disk generously — model weights, LoRAs and outputs add up, and disk bills separately per GB-month.

  • 02

    Choose ComfyUI or A1111

    Both ship with the common nodes and dependencies installed. After creating the instance, wait for the image pull; only storage bills during that window.

  • 03

    Open the WebUI in your browser

    The instance page hands you the WebUI address and it opens straight into the familiar interface. Download models from within the UI or wget them from the web terminal.

06 — FAQ

AI art and video generation

What do 100 SDXL images cost?

At 1024px an RTX 4090 needs roughly fifteen minutes for 100 images, which is under twenty cents at $0.540 per hour. At that scale cost is essentially noise — the thing to remember is to destroy the instance afterwards, because disk keeps billing while it is merely stopped.

How do I get my models and LoRAs onto the machine?

Three ways: paste a Civitai or Hugging Face link into the WebUI, wget from the web terminal, or scp from your machine. Prefer the first two for large files — they use the host's public bandwidth, which is far faster than uploading from home.

Does FLUX really need 24GB?

For the original FLUX.1 dev, yes. A 12GB card can run GGUF quantised builds, but detail and quality visibly suffer. Since a 24GB card is $0.540 per hour, there is little reason to settle for a quantised build to save that.

Can the platform see my generated images?

The instance is a container you hold exclusively; we neither read nor back up its disk. Destroying it releases the disk. Note the corollary: there is no automatic backup, so download anything that matters.

Can I run several instances in parallel?

Yes — there is no instance cap beyond your balance. Three to five parallel instances is common for batch style testing. Total cost matches one instance running five times as long, but you get results four times sooner.

Related solutions

Other ways to use it

Same compute network — swap the image and it becomes a different production line.

Start building on NexGPU

Enterprise R&D team or solo developer — either way, your first job can be running in minutes.

Sign up freeTalk to sales

Sign up to browse live network pricing. No payment method required.