Krea 2 Locally: Setup, VRAM, and the License Catch

Yes — Krea 2 runs locally. Krea released open weights on June 22, 2026: a 12.9B diffusion transformer in two checkpoints, Raw (undistilled) and Turbo (8-step distilled). ComfyUI supports it natively since v0.26.0. The catch: the Krea 2 Community License only allows free commercial use under $1M annual revenue and 50 seats.

That last sentence is the part most coverage skips, and it changes who should build on this model. This guide walks through what actually shipped, where to download it, the exact ComfyUI file layout from the official docs, honest VRAM math (Krea has not published official minimums — we show the arithmetic instead), and the license terms spelled out plainly. Then we put it next to FLUX.2 klein, the other current-gen local image model you are probably deciding between.

---

What Krea Actually Released {#what-is-krea-2}

On June 22, 2026, Krea — until then known as a hosted creative tool — published open weights for its image model. The release is a 12.9B-parameter diffusion transformer, shipped as two separate checkpoints on Hugging Face:

One day later, on June 23, 2026, ComfyUI shipped v0.26.0 with native Krea support, per the official ComfyUI changelog. No custom nodes, no community wrapper repos — it works out of the box on an updated install.

The license is the asterisk on all of this. It is a custom "Krea 2 Community License", not Apache 2.0 and not MIT. It is genuinely free for most individuals, but it has hard commercial cutoffs we cover below.

---

Raw vs Turbo {#raw-vs-turbo}

Same architecture, two different jobs.

Krea 2 Turbo is the one to run day to day. It is distilled to generate in 8 steps — Krea's official example code uses num_inference_steps=8 with guidance_scale=0.0. Krea claims Turbo produces 2K images in roughly 2 seconds "on consumer hardware." That is a vendor claim — they do not name the GPU, and we have not independently timed it — but 8-step distillation making generation feel instant is consistent with how other distilled models behave.

Krea 2 Raw is the undistilled base. In ComfyUI it samples at the full 52 steps, so it is slow by comparison — but distilled models are poor targets for training, which makes Raw the checkpoint for fine-tuning and LoRA work. If you are building a style LoRA on Krea 2, you train against Raw. Krea's own style LoRAs load into the Turbo workflow's LoRA slot, and community fine-tunes are expected to work the same way.

Simple rule: generate on Turbo, train on Raw.

---

Downloading the Weights {#download}

Both checkpoints live on Hugging Face:

• Turbo: huggingface.co/krea/Krea-2-Turbo
• Raw: huggingface.co/krea/Krea-2-Raw

For ComfyUI you need three files total (exact names in the next section): the diffusion model, a Qwen3VL-4B text encoder, and the qwen_image_vae VAE — all linked from the official ComfyUI docs for Krea 2.

On precision: the ComfyUI docs recommend the FP8 checkpoint for most users, with BF16, NVFP4, and MXFP8 variants "for users with higher-end hardware." The official Hugging Face card lists F32 and BF16 tensor types. We found no official GGUF release at the time of writing — if you see one, it is a community conversion; check who uploaded it before trusting it.

---

ComfyUI Setup (v0.26+) {#comfyui-setup}

First, update ComfyUI to v0.26.0 or later. Support is native — if your install predates June 23, 2026, the Krea nodes won't exist yet and you'll hit missing-node errors when you load the workflow.

File placement, straight from the official ComfyUI docs:

Then load the built-in Krea 2 template workflow. The defaults — 8 steps, prompt enhancement on, no LoRA — produce a good image without touching anything. For CFG, ComfyUI's docs do not specify a value, but Krea's own example code pins guidance_scale=0.0 for Turbo, which is normal for distilled models: guidance is baked in during distillation, so you leave it off at inference.

Running Raw instead? Swap the diffusion model and expect 52 steps per image.

New to ComfyUI entirely? Start with our ComfyUI complete guide first — the Krea workflow will make much more sense.

---

VRAM: The Honest Math {#vram}

Here is the uncomfortable truth: neither the Hugging Face card nor the ComfyUI docs publish VRAM requirements. Anyone quoting you an exact official minimum is making it up. What we can do is arithmetic — clearly labeled as estimates:

On top of the diffusion weights you are also loading the Qwen3VL-4B text encoder (FP8 variant, a few GB), the VAE, and activation memory during sampling. ComfyUI loads and unloads models in stages, which is why a 16GB card is a plausible floor for the FP8 checkpoint — tight, with offloading doing real work — while 24GB is the comfortable tier. The NVFP4 variant exists for hardware that supports it and should land smaller than FP8, but nobody publishes a number, so we will not invent one.

For context on what image models actually fit per GPU, our FLUX VRAM breakdown by card covers the same memory logic in detail, and if you are shopping in the 16GB class, see the best image model picks for 16GB VRAM.

---

The License Catch {#license}

This is the section to read before you build anything commercial on Krea 2.

Krea 2 ships under a custom "Krea 2 Community License." It is not Apache 2.0. It is not MIT. The headline terms:

1. Free commercial use only under $1M annual revenue AND under 50 seats. Both conditions. Cross either line — revenue or headcount — and you need an enterprise license from Krea.
2. Mandatory safety guardrails. The Hugging Face card states it directly: "Deployers are required to implement content filtering measures or equivalent review processes to prevent the generation or distribution of unlawful or policy-violating content." That is a real compliance obligation if you deploy Krea 2 in a product, not boilerplate.

What this means in practice:

• Hobbyists, freelancers, small studios: you are fine. Under $1M and under 50 seats describes almost everyone reading a local-AI setup guide.
• Agencies, funded startups, growing teams: read the actual license text before you standardize on this model. $1M in revenue arrives faster than most teams expect, and "seats" is defined by the license, not by your assumptions.
• Anyone shipping a product with Krea 2 inside: the content-filtering requirement applies to you regardless of size.

Compare that to FLUX.2 klein's Apache 2.0 — no revenue cap, no seat cap, no mandated guardrails. The license difference alone decides the choice for some teams, which brings us to the head-to-head.

---

Krea 2 vs FLUX.2 klein {#krea-2-vs-flux}

These are the two current-gen open-weights image models most people are choosing between for local use. They occupy opposite corners:

The short version:

• Pick FLUX.2 klein if you have 8-12GB of VRAM, want a license nobody has to think about, or are shipping commercially without wanting a legal review. A 2.6GB Q4 GGUF runs on hardware Krea 2 cannot touch. Setup details in our FLUX.2 local guide.
• Pick Krea 2 if you have 16-24GB+, qualify under the Community License, and want the larger model. At 12.9B vs 4B it is more than 3x the parameter count — though we have deliberately not claimed a quality verdict here, because we have not run a controlled comparison yet.

For the full field beyond these two — including Stable Diffusion 4 — see our local image model comparison.

---

Verdict {#verdict}

Krea 2 is a real local release, not a marketing gesture: proper open weights on Hugging Face, two sensibly split checkpoints, and native ComfyUI support within a day of launch. If you have a 16GB card, the FP8 Turbo checkpoint is worth an afternoon. If you have 24GB, it is comfortable.

Two honest caveats. First, the VRAM story is entirely community arithmetic right now — Krea publishes no official minimums, so budget conservatively. Second, and more importantly, the Community License is doing quiet work. Under $1M revenue and under 50 seats it costs you nothing; the moment your business grows past either line, this model has a price tag where FLUX.2 klein's Apache 2.0 does not. Run Krea 2 locally, enjoy it — but if it is going inside a product, read the license before the model becomes load-bearing.

---

Sources {#sources}

• Krea 2 Turbo on Hugging Face — official model card: tensor types, example inference code, license
• Krea 2 Technical Report — Krea's official release post and technical details
• ComfyUI v0.26.0 changelog and official Krea 2 docs — native support, file placement, default workflow settings

---

FAQ {#faq}