Krea 2 is one of the most capable open-weights image models of 2026 — but “open weights” doesn’t mean “easy to run.” Between the gated download, the ComfyUI wiring, and the VRAM it wants, most people who search for “how to run Krea 2” actually just want to use it without turning their laptop into a render farm. This guide covers both: what it takes to run Krea 2 locally, and how to generate with it online in a couple of clicks.
What is Krea 2?
Krea 2 is a text-to-image diffusion model released as open weights by Krea under the Krea 2 Community License. It ships in two flavours:
- Krea 2 Turbo — a distilled, fast variant. It produces a full image in roughly 8–12 steps at CFG 1, which makes it ideal for interactive generation.
- Krea 2 Raw — the un-distilled model. It runs at 25–30 steps with a higher CFG (~4), giving a small quality edge at the cost of being several times slower.
Its most distinctive trait is the text encoder: Krea 2 uses a Qwen3-VL language model rather than a classic CLIP encoder. In plain terms, Krea 2 reads your prompt like a sentence — it understands natural prose far better than comma-separated tags. (More on that in our prompting guide.)
Why running Krea 2 locally is hard
If you want to self-host, here’s the honest checklist:
- The weights are gated. The base model lives on Hugging Face at
krea/Krea-2-Turboand requires accepting the license before you can download. (Small gotcha: use the exact capitalisation — the lower-case form can 307-redirect, and not every download tool follows that.) - The full-precision model is big. The bf16 diffusion weights are about 26 GB on their own, before you add the ~5 GB text encoder and the VAE.
- It’s a multi-component ComfyUI pipeline. You need the diffusion model, the Qwen3-VL text encoder, and the matching VAE — and the CLIP loader has to be set to the
krea2type or the prompt won’t encode correctly. - Newer GPUs need newer CUDA. Blackwell-generation cards (RTX 50-series) require a PyTorch build on CUDA 12.8+, or the GPU simply won’t be detected.
You can shrink it with quantisation — a 4-bit nvfp4 build fits in about 7 GB and runs on an 8 GB card — but you trade away fine detail, and generation is noticeably slower than a hosted GPU. For a full breakdown of what hardware you need, see our Krea 2 VRAM & system requirements guide.
The fast path: run Krea 2 online
If you’d rather skip all of that, AIFLUX runs Krea 2 Turbo for you in the cloud. You get:
- No GPU and no install — it runs in any modern browser.
- Natural-language prompting — write a real sentence; the model’s LLM text encoder does the rest.
- Free credits to start, then simple pay-as-you-go. No subscription.
- Your outputs are yours — under the Krea 2 license, Krea claims no ownership of generated images.
👉 Open the Krea 2 generator on AIFLUX and type your first prompt.
Krea 2 online vs. local — which should you pick?
| Run online (AIFLUX) | Run locally | |
|---|---|---|
| Setup time | None | Hours (ComfyUI + weights + CUDA) |
| Hardware | Any browser | 8 GB VRAM (quantised) to 24 GB+ (bf16) |
| Speed | Cloud GPU | Depends on your card |
| Best for | Getting results now | Tinkering with the full pipeline |
Local is great if you want to hack on the graph and control every node. Online is the right call if you want good Krea 2 images today without buying a GPU or debugging CUDA.
Frequently asked questions
Is Krea 2 free? The model weights are open under a community license, and you can run them yourself for free if you have the hardware. Running online on AIFLUX starts with free credits, then pay-as-you-go.
Do I need to download anything to use Krea 2? No — on AIFLUX you generate in the browser. You only need downloads if you self-host in ComfyUI.
What’s the difference between Krea 2 Turbo and Raw? Turbo is fast (8–12 steps, CFG 1); Raw is slower but slightly higher quality (25–30 steps, CFG ~4). AIFLUX serves the Turbo variant for interactive speed.
Can I use Krea 2 images commercially? The Krea 2 Community License permits commercial use of outputs (subject to its terms, including a company-revenue threshold and content-safety obligations). You own the images you generate. Always check the current license text for your situation.
Ready to try it? Generate with Krea 2 on AIFLUX →