📊 See Krea 2 repo rankings

📱 Get the AI Rank app

Krea 2 is Krea's own image model, trained from scratch. Its open release ships as two checkpoints: RAW, the undistilled base model, and Turbo, an 8-step distilled model for everyday generation. Both are 12B-parameter diffusion transformers. Hugging Face hosts the weights behind an access form, and the official model card gives June 22, 2026 as the release date. Krea's short version of how to use them: "TRAIN on Raw and RUN on Turbo."

Krea and Comfy document three local routes: the official inference script, Diffusers and ComfyUI. LoRA training happens on RAW, and commercial use is governed by the Krea 2 Community License, which has a revenue cap.

What Ships in the Open Release

The krea-ai/krea-2 repository holds the inference code under Apache-2.0. The weights carry a separate license, covered below. According to Krea's release page, the architecture is a 12B dense DiT that uses the Qwen Image VAE and a Qwen3-VL text encoder with multi-layer feature aggregation.

Item Krea 2 RAW Krea 2 Turbo
Hugging Face repo krea/Krea-2-Raw krea/Krea-2-Turbo
Single-file checkpoint raw.safetensors, about 26.3 GB turbo.safetensors, about 26.3 GB
Training Pretrained, no distillation, no post-training Distilled from a post-trained checkpoint
Recommended steps / CFG 52 steps, CFG 3.5 8 steps, CFG 0
Resolution Trained up to 1K 1K to 2K
Intended job Fine-tuning, LoRA training, research Fast text-to-image

Each Hugging Face repo also includes a Diffusers folder layout (transformer, text encoder, VAE, tokenizer, scheduler) alongside the single file. On October 19, 2026, AI Rank listed the GitHub repo at 929 stars and #6,061 overall, filed under Vision models.

Getting Access to the Weights

Both Hugging Face repos are gated. To request access, you sign in, give your name, email and company, and tick "I agree to the Krea 2 Community License." Once approved, your token can download the files. The ComfyUI-repackaged files in Comfy-Org/Krea-2 were not gated as of October 19, 2026, but the same community license applies to them.

Running Krea 2 With the Official Script

The repository uses uv. Its pyproject.toml asks for Python 3.12 or newer, PyTorch 2.9 or newer and transformers 4.57.1, and pulls PyTorch from the CUDA 12.8 index.

git clone https://github.com/krea-ai/krea-2
cd krea-2
uv sync
export OSS_RAW=/path/to/raw.safetensors
export OSS_TURBO=/path/to/turbo.safetensors

These are the README's recommended commands, RAW first (full sampler with classifier-free guidance), then Turbo (8 steps, CFG off, 2K output):

uv run inference.py "a fox walking in the snow" \
    --checkpoint oss_raw --steps 52 --cfg 3.5

uv run inference.py "a fox walking in the snow" \
    --checkpoint oss_turbo --steps 8 --cfg 0.0 --mu 1.15 --width 2048 --height 2048

Pass those flags explicitly. If you leave them out, the script defaults to 28 steps, CFG 4.5 and the oss_raw checkpoint, which matches neither recommended setup. Other flags include --seed (image i uses seed + i), --num-images and --output. Width and height are padded up to a multiple of 16.

Krea's prompting guide recommends natural-language prompts, with long detailed prompts giving the best results. For text in the image, put the words in quotes. The repo also includes expansion.txt, a system prompt you can give any LLM to expand a short idea into a longer prompt.

Diffusers and SGLang

The model cards show a Krea2Pipeline that needs Diffusers installed from source:

pip install git+https://github.com/huggingface/diffusers.git

import torch
from diffusers import Krea2Pipeline

pipe = Krea2Pipeline.from_pretrained("krea/Krea-2-Turbo", torch_dtype=torch.bfloat16).to("cuda")
image = pipe("a fox in the snow", num_inference_steps=8, guidance_scale=0.0).images[0]
image.save("krea2.png")

For RAW, load krea/Krea-2-Raw and use num_inference_steps=52, guidance_scale=3.5. SGLang, also installed from source, has a one-line CLI: sglang generate --model-path krea/Krea-2-Turbo --prompt "..." --num-inference-steps 8 --height 1024 --width 1024 --save-output.

Choosing Between RAW and Turbo

Pick Turbo for making images. It is distilled from a post-trained RL checkpoint, so Comfy's launch post describes its output as higher quality than RAW's but with less variation between images. It runs in 8 steps and goes up to 2K.

Pick RAW for training. It is the checkpoint before any distillation or post-training, more varied and easier to push toward a style or subject. The Hugging Face card says RAW is "not recommended for inference use," and the README gives 1K as its training resolution. A LoRA trained on RAW is meant to be applied on Turbo.

Using Krea 2 in ComfyUI

Comfy announced native support on June 23, 2026 and told users to update to ComfyUI 0.26.0. In an updated install, search "Krea-2" in the Template Library, or download image_krea2_turbo_t2i.json from Comfy's workflow templates repo. Comfy's docs recommend these files from Comfy-Org/Krea-2 for most local users:

ComfyUI/models/
├── diffusion_models/krea2_turbo_fp8_scaled.safetensors   (about 13.1 GB)
├── text_encoders/qwen3vl_4b_fp8_scaled.safetensors
├── vae/qwen_image_vae.safetensors
└── loras/   (optional style LoRAs)

The same repo also has BF16, MXFP8, Int8 and NVFP4 builds. The NVFP4 Turbo file is the smallest at about 7.7 GB. The template's defaults are 8 steps, LLM prompt enhancement on and no LoRA. Set megapixels to 2.0 in the ResolutionSelector for 2K output. A second template, "Krea-2 Style Reference," uses the Int8 Turbo file and takes one or two reference images.

ComfyUI also has Krea 2 partner nodes that call Krea's hosted API. Those are a separate, cloud-billed product, not the local open weights described here.

Krea released nine style LoRAs. Comfy lists a recommended strength of 1.0 for each, and every LoRA has a trigger word:

LoRA Trigger word
darkbrush monochrome ink wash style
dotmatrix monochrome stippling style
kidsdrawing naive expressive sketch style
neondrip textured abstract style
rainywindow rainy window style
retroanime purple retro anime style
softwatercolor art deco watercolor style
sunsetblur ethereal motion blur style
vintagetarot vintage tarot style

→ More open models: Open-source AI projects on AI Rank

📱 Track them in the AI Rank app

Training a LoRA

Krea's README is direct about the setup: train the LoRA on RAW, run it on Turbo. It names four tools:

The README doesn't publish training hyperparameters or hardware requirements. Check each tool's own Krea 2 docs for those. The nine official style LoRAs show what a finished LoRA looks like: one file per style, loaded on Turbo with its trigger word.

License and Commercial Use

The weights use the Krea 2 Community License (v.1, June 22, 2026), not Apache-2.0. The terms that matter most:

  • Commercial use has a revenue cap. It is allowed only if you, together with every affiliate under common ownership or control, earn less than $1,000,000 USD a year, counted over the trailing twelve months from all sources. At or above that, you need an Enterprise License from Krea before any commercial use. Contact: opensource@krea.ai.
  • Outputs count. The license defines commercial use as covering the model, derivatives and outputs, so selling generated images falls under the same cap.
  • You own your outputs. Krea claims no ownership of generated images.
  • Distribution rules. If you share the model or a fine-tune, include the license, start the model name with "Krea" (the license's example is "Krea 2 [Your Model Name]"), ship a Notice file with the required attribution, and state that you modified it.
  • Content filtering is mandatory. Section 4.2 requires deployers to put content filters in place, such as prompt or output filtering, abuse detection or human review. Krea's safety doc says failing to do so breaches the license.
  • AI disclosure. Where law, regulation or platform policy requires it, you must disclose that outputs are AI-generated.

FAQ

Is Krea 2 open source? The inference code is Apache-2.0. The weights are open-weight under the Krea 2 Community License, which caps commercial use by revenue and adds distribution and filtering rules.

What are the recommended settings? RAW: 52 steps, CFG 3.5. Turbo: 8 steps, CFG 0, and --mu 1.15 in the official script.

How much VRAM does Krea 2 need? Neither Krea nor Comfy publishes a figure. On disk, the BF16 checkpoint is about 26.3 GB, and Comfy's FP8 Turbo file is about 13.1 GB.

Can I use RAW to generate final images? It runs, but the Hugging Face card doesn't recommend it for inference. Use Turbo.

Is the Krea 2 on krea.ai the same model? The model card lists both an open-weight release and Krea-hosted integrations. ComfyUI's Krea 2 partner nodes go through Krea's API and are billed separately.

Keep exploring: compare Krea 2 with other open models in the open-source rankings on AI Rank.

Or 📱 get the AI Rank app to follow new releases.

Sources

Prepared with automated editorial assistance and checked on October 19, 2026 against the krea-ai/krea-2 repository, the Hugging Face model cards and file listings, Krea's release and license pages, ComfyUI documentation and AI Rank data. Krea 2 was not downloaded or run for this article.