The single most common question about Krea 2 is “will it run on my GPU?” The honest answer is: it depends heavily on which build you use. Krea 2 ships in full precision and in several quantised builds, and the VRAM gap between them is enormous. Here’s what you actually need.
Model sizes at a glance
Krea 2 is not one file — it’s a pipeline. Budget for all three components:
| Component | Purpose | Size (bf16) |
|---|---|---|
| Diffusion model | The image generator itself | ~26 GB |
| Text encoder (Qwen3-VL) | Reads your prompt (it’s an LLM) | ~5 GB |
| VAE | Decodes latents into pixels | small (~0.3 GB) |
That’s why running the full bf16 model realistically wants a 24 GB GPU (e.g. an RTX 4090/5090-class card) — and even then you should avoid loading two full-precision models at once.
The quantisation ladder
You don’t have to run full precision. Quantised builds trade a little quality for a lot less VRAM:
- bf16 — best quality, ~26 GB model, needs a big GPU. This is what cloud services run.
- nvfp4 (4-bit) — about 7 GB, the smallest practical build. It fits on an 8 GB card. Expect slightly softer fine detail versus bf16, and slower generation.
A realistic rule of thumb on an 8 GB card running the nvfp4 build: a 1024×1024 image lands in roughly a minute without extra LoRAs, and VRAM sits close to the ceiling. Push past ~1.5 megapixels and you’ll want tiled VAE decoding or a low-VRAM mode to avoid running out of memory.
GPU & CUDA requirements (don’t skip this)
Two hardware gotchas trip up most first-time setups:
- Blackwell needs CUDA 12.8+. RTX 50-series cards (compute capability sm_120) require a PyTorch build compiled for CUDA 12.8 or newer. Install torch with the correct CUDA index URL — a generic
pip install torchcan pull a build that doesn’t see your GPU at all. - Don’t blind-upgrade torch. If Krea 2 is working and your GPU suddenly “disappears,” a torch update that dropped you off the CUDA 12.8 wheels is the usual culprit.
For 30-series and 40-series cards, standard current CUDA builds are fine; the Blackwell note is specific to the newest generation.
Local vs. cloud: the practical trade-off
Local generation is great for experimentation, but the quantised build’s quality and speed are a step below a hosted bf16 GPU. Many people use a small local build to draft prompts and composition, then run the final, high-quality images in the cloud.
If you don’t have a capable GPU — or you just don’t want to manage CUDA versions and 26 GB downloads — you can run Krea 2 Turbo online on AIFLUX. It runs the bf16 model on cloud GPUs, so there are no VRAM requirements on your side at all: it works on a laptop, a Chromebook, or a phone.
👉 Generate with Krea 2 online — no GPU needed
Frequently asked questions
How much VRAM do I need for Krea 2? For the full bf16 model, plan on a 24 GB GPU. For the 4-bit nvfp4 build, 8 GB is enough. Online, you need none.
Can Krea 2 run on 8 GB of VRAM? Yes, using the nvfp4 (4-bit) build, at ~1024×1024. For larger images use tiled VAE decoding or a low-VRAM flag.
Does Krea 2 work on RTX 5070 / 5080 / 5090? Yes, but you must install PyTorch built for CUDA 12.8+ (Blackwell / sm_120). Without it, the card won’t be detected.
Do I need CUDA 12.8 for Krea 2? Only on Blackwell (RTX 50-series) GPUs. Older cards run on current standard CUDA builds.
Is there a way to use Krea 2 without a good GPU? Yes — run it online on AIFLUX, which hosts the model on cloud GPUs.
Skip the hardware math. Run Krea 2 Turbo in your browser →