Krea 2 Locally: Open Weights, VRAM and What the Licence Allows

Krea 2 is the open image model from Krea, the AI image company, released on 18 June 2026 in two versions: Raw, an undistilled base for training, and Turbo, an 8-step model for generating. Comfy-Org’s repackaged files for ComfyUI have been downloaded about 3.4 million times in the past month, making it one of the most-used open image models of 2026.

It is also the strictest of this year’s open image models on commercial use, even though Krea describes its licence as permissive. This guide covers what to download, how much memory it takes, the settings Krea recommends, and exactly what the Krea 2 Community License allows. BFWR has not run Krea 2; memory figures are our calculated estimates, labelled as such, and everything else is from Krea’s and Comfy-Org’s own files and documentation.

Quick answer

• Two checkpoints: Turbo for making images (8 steps, no guidance), Raw for training LoRAs. Krea’s advice is “TRAIN on Raw and RUN on Turbo”.
• ComfyUI: native support since June, with official local templates.
• Download: about 18.6GB for the files the standard template uses.
• Memory: about 16GB at 1 megapixel with 8-bit weights and the text encoder offloaded (estimate). It is the largest image model of this year’s new releases.
• Licence: commercial use only if your organisation’s revenue is under US$1 million a year, with mandatory content filtering.

What Krea 2 is

Krea’s GitHub repository describes it as “an image model trained from scratch focused on creative and stylistic exploration”. Its technical report, published on 23 June 2026, describes a single-stream diffusion transformer that uses Qwen3-VL as its text encoder. The weights put it at 12.82 billion parameters, a 26.3GB transformer at BF16, with Qwen3-VL 4B as the text encoder (the same encoder FLUX.3 Image Klein uses) and the Qwen-Image VAE.

Krea’s README is direct about how they fit together: “You train LoRAs on RAW and apply them on Turbo, and the LoRAs trained on RAW will work well on Turbo.” It recommends diffusers, Ostris AI Toolkit, Kohya’s musubi tuner and fal for training.

Krea makes strong claims about quality, calling Krea 2 “the most aesthetic open-source image model available”, and its README calls it the top text-to-image model from an independent lab on Artificial Analysis’s leaderboard. Its own technical report is more modest, placing it “2nd place among models from independent labs”. Either way, those are the company’s claims. Note also that Krea 3, added to ComfyUI on 15 October 2026, is a different thing: Comfy-Org’s templates list it only as API nodes, which call Krea’s servers. Krea 2 is the open model you run yourself.

Running it in ComfyUI

ComfyUI added Krea 2 support in June 2026. Its official templates list ComfyUI 0.24.1 as the minimum, so any recent version has it. Three local templates are in the template browser:

• Krea-2: Text to Image, using the FP8 Turbo model
• Krea-2 Int8: Text to Image, using the INT8 Turbo model
• Krea-2 Int8: Image Style Reference, which adds a style-reference LoRA

Both text-to-image templates use 8 steps, CFG 1 (no guidance), the euler sampler and the simple scheduler, at 1024×1024. They include an optional prompt-enhancement step that rewrites short prompts using the Qwen3-VL text encoder, and an optional style LoRA, both switched from inside the template. The text encoder loads through Load CLIP with the type set to krea2.

Files and download size

Comfy-Org’s repackaged files are not gated. Krea’s own repositories are, and require accepting the licence on Hugging Face. Sizes from the Comfy-Org listing:

The standard template’s set, FP8 Turbo with the FP8 text encoder and the VAE, comes to about 18.6GB. NVFP4 is NVIDIA’s 4-bit format for its Blackwell generation, so the 7.67GB file is mainly useful on RTX 50-series cards. If you already run FLUX.3 Image Klein, the Qwen3-VL 4B text encoder is the same model, though Comfy-Org packages it separately for each.

How much memory it needs

Krea publishes no VRAM figure. These are estimates from the formula behind our image generation VRAM calculator: weights at each precision plus an activation allowance per megapixel, borrowed from FLUX.1, a transformer of similar size. Treat them as planning figures with headroom.

In practice, Krea 2 at 8-bit is a 24GB-card model, with a 16GB card workable at 1 megapixel once ComfyUI moves the text encoder off the GPU. On 12GB you are down to 4-bit with heavy offloading, which costs speed. For comparison on the same formula, at 1 megapixel with 8-bit weights and the encoder offloaded, FLUX.3 Image Klein needs about 8.7GB and Qwen-Image 2.1 about 10.8GB. Krea 2’s transformer is more than twice the size of FLUX.3 Klein’s, which is where the difference comes from.

Our model page, Krea 2 VRAM requirements, shows every precision and offload mode.

The licence: Krea 2 Community License

Krea’s README says both checkpoints are under its “community license with permissive use”. The licence text is more restrictive than that phrase suggests, especially for businesses. It is published, ungated, in Krea’s GitHub repository as the Krea 2 Community License Agreement, version 1, dated 22 June 2026. What follows is a reading of that text, not legal advice.

The revenue limit. Section 2.3 permits commercial use “only if you (including all affiliated entities under common ownership or control) have total company-wide annual revenue of less than one million United States dollars ($1,000,000 USD)”, measured over a trailing twelve months. If you reach that threshold at any point, you “must immediately cease Commercial Use and contact Krea”.

What counts as commercial. The definition is broad: any use of “the Krea Model, Derivatives, or Outputs in connection with any business, trade, or commercial activity”. That includes the images. This is the important difference from FLUX.3 Image Klein’s licence, which sets a higher US$5M threshold and applies it to using the images commercially (see our FLUX.3 Image Klein guide).

Rules that apply to everyone, commercial or not:

• Content filtering is mandatory. “You must implement reasonable and appropriate Content Filter measures” to prevent harmful or unlawful content. The licence gives open-source classifiers and manual review as examples.
• AI disclosure is required where law, regulation or platform policy demands it.
• Naming: if you distribute a model built on Krea 2, its name must start with Krea (the licence’s example is “Krea 2 [Your Model Name]”), and you must keep a notice file and include a copy of the licence.
• It can be withdrawn. The grant is “revocable”, and section 9.2 lets Krea end the agreement “for any reason upon thirty (30) days’ prior written or electronic notice to you.” That is worth weighing before you build a product on it.

What is generous: section 5.3 says “You own all Outputs you generate”, and section 5.2 gives you ownership of derivatives you create, subject to Krea’s rights in the base model. For individuals and small studios under the threshold, that is a usable licence. For anyone above it, commercial use needs an enterprise licence from Krea.

If you need an image model with no revenue condition, MOSS-Image and FLUX.2 [klein] 4B are both Apache 2.0. Our MOSS-Image guide covers the first.

LoRAs

Krea released nine official style LoRAs alongside the model (dark brush, dot matrix, kids’ drawing, neon drip, rainy window, retro anime, soft watercolour, sunset blur and vintage tarot), each about 0.47GB, and Comfy-Org packages a style-reference LoRA from Ostris. Each Krea style LoRA has a trigger phrase, listed on Comfy-Org’s model card (for example “vintage tarot style” or “monochrome ink wash style”), at a recommended strength of 1.0. In the official template you pick the LoRA from a list and the trigger phrase is added to your prompt for you. The rule from our guide to matching LoRAs to the base model applies: Krea 2 LoRAs are built for Krea 2’s transformer and will not work on FLUX or Qwen-Image, even though the text encoder is shared. Under the licence, a LoRA is a derivative, so distributing one brings the naming and notice rules with it.

FAQ

Is Krea 2 free?

Free to download and use non-commercially. Commercial use is free only for organisations with under US$1M in annual revenue.

Turbo or Raw?

Turbo to make images: 8 steps, no guidance, 1K to 2K. Raw to train LoRAs, which Krea says then work on Turbo.

Can I run Krea 2 on a 12GB graphics card?

Only at 4-bit with offloading, by our estimate about 10.7GB at 1 megapixel, which leaves little headroom. 16GB is a more realistic minimum for 8-bit, and 24GB comfortable.

Is Krea 3 open source?

Not as of 18 October 2026. In ComfyUI it appears only as API templates that call Krea’s service.

Related guides and tools

• Qwen-Image 2.1 in ComfyUI, whose VAE Krea 2 uses
• ComfyUI templates: local versus API
• Best local image generation models in 2026, every model compared on memory and licence
• ComfyUI hub and image model directory
• Stable Diffusion Prompts

Sources

• krea-ai/krea-2 on GitHub: README, recommended settings and the Krea 2 Community License
• Krea 2 technical report (23 June 2026)
• Comfy-Org/Krea-2 repackaged files, and krea/Krea-2-Turbo (parameter count)
• Comfy-Org workflow templates (Krea 2 and Krea 3 entries)

Updated 18 October 2026. File sizes, template settings and licence text were read on that date. Memory figures are estimates, not measurements.