Krea Opens Krea 2's 12B Image Model So Anyone Can Train and Deploy

Krea just dropped the open weights of Krea 2, its flagship image generation model, in two flavors: Krea 2 Raw and Krea 2 Turbo. The headline pitch is simple: train on Raw, generate with Turbo. Together, they hand the community a 12-billion-parameter aesthetic image model that runs on consumer hardware and ships with a commercially friendly license.

Two weights, one strategy

The release is not a single model dump. Krea is releasing two distinct checkpoints that serve very different roles in a workflow:

• Krea 2 Raw is a mid-training checkpoint, snapshotted before any distillation, fine-tuning, or post-training. Think of it as the model in its most malleable state, before Krea baked in their aesthetic preferences. It is explicitly not recommended for direct inference, but it is the ideal base for training LoRAs (low-rank adapters, a lightweight fine-tuning technique that adds small trainable matrices to a frozen model) and full fine-tunes.
• Krea 2 Turbo is the polished, production-ready version. It has been distilled (compressed to run in far fewer steps) and post-trained for aesthetic quality. Turbo is an 8-step distilled checkpoint built for fast, high-quality generation at 1K to 2K resolution. It runs with guidance_scale=0.0, meaning no classifier-free guidance overhead.

The intended workflow is: train your LoRA on Raw, then run it on Turbo for inference. LoRAs trained on the Raw base model transfer well to Turbo for inference. Krea also released a collection of in-house LoRAs trained on Raw and meant to be used with Turbo, so you can see this pattern in practice immediately.

What is under the hood

Krea 2 is a latent-diffusion image model trained from scratch with an emphasis on aesthetics and stylistic control. The architecture is a single-stream multimodal diffusion transformer. The key architectural choices are worth knowing:

• The transformer is a single-stream DiT with 12.9B parameters, 28 blocks at width 6144, using grouped-query attention, a learned output gate, per-head QK normalization, and a 3-axis rotary embedding.
• The text encoder is Qwen/Qwen3-VL-4B-Instruct, tapped at twelve intermediate layers for text-only conditioning. Rather than using just the final output of the text encoder, Krea pulls features from 12 different layers and fuses them inside the transformer, giving the model a richer, multi-scale understanding of your prompt.
• The VAE is the Qwen-Image autoencoder (AutoencoderKLQwenImage, f8, 16 latent channels), with a flow-matching sampler using a resolution-aware timestep shift.

Krea 2 Open-Source uses a Qwen Image VAE, a 12B dense DiT backbone, and a Qwen3-VL text encoder with multi-layer feature aggregation. The choice of Qwen3-VL as the text encoder is notable: it is a vision-language model repurposed purely for text conditioning, bringing strong language understanding to image generation without the typical CLIP bottleneck.

Getting started in two lines

Both models are available on Hugging Face and integrate with diffusers. You need to install diffusers from source to get the Krea2Pipeline:

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

For Turbo (fast inference):

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 (fine-tuning base):

pipe = Krea2Pipeline.from_pretrained("krea/Krea-2-Raw", torch_dtype=torch.bfloat16).to("cuda")
image = pipe("a fox in the snow", num_inference_steps=52, guidance_scale=3.5).images[0]

Notice the difference: Turbo runs in 8 steps with no guidance, Raw needs ~52 steps with CFG. Turbo also supports SGLang for CLI-based inference, which is useful for batched server deployments.

Where it shines, where it does not

Krea 2 Open-Source brings Krea 2, their aesthetic-first model family, into the open with open weights on Hugging Face. The model is explicitly designed for creative and visual production work. Krea highlights several strong use-cases:

• Exploring materials, lighting, interiors, facades, and early design directions with an open image model built for aesthetic control.
• Generating campaign directions, editorial frames, product concepts, and visual territories before a shoot or production pass.
• Moving from rough concepts to polished visual systems across brand, poster, packaging, interface, and moodboard exploration.

On the flip side, the model cards are explicit that Krea 2 Raw is not recommended for inference use. And like all current image models, prompt following may be influenced by prompt style, specificity, language, and phrasing , so expect to iterate. The model is also not designed to provide factual information, and text rendering in images remains a known weak point for diffusion models generally.

The licensing math

The license is the Krea 2 Community License, and the key number is $1M ARR. Commercial use under this agreement is permitted only if you have total company-wide annual revenue of less than one million United States dollars, calculated on a trailing twelve-month basis and including all revenue from all sources. Above that threshold, you need an enterprise license. For individuals, researchers, and early-stage startups, this is effectively free commercial use.

A few other license details worth noting:

• You own the outputs you generate. Krea claims no copyright over generated images.
• If you distribute derivatives, the model name must start with "Krea" (e.g., "Krea 2 YourModelName").
• Deployers must implement content filtering, with the license listing specific open-source options like Falconsai/nsfw_image_detection and NudeNet as examples.

The bigger picture

This release follows Krea's earlier open-source move with FLUX.1 Krea, where they post-trained Black Forest Labs' foundation model with their aesthetic preferences. With Krea 2, they trained the foundation model themselves from scratch, which means they can now release the actual mid-training checkpoint rather than a fine-tune on someone else's base. That is the key difference: Raw is a genuine pre-post-training base, not a fine-tune of an existing open model.

The open-weights image model space is heating up fast. Krea 2 enters alongside FLUX.1 variants and Stable Diffusion 3.5, but its differentiator is the explicit Raw/Turbo split designed around the LoRA training workflow. The ecosystem is already responding: the architecture is already supported by partners including Hugging Face, SGLang, FAL, ComfyUI, Nous Research, Replicate, Cloudflare, Together AI, Google GCP, and Amazon AWS. Both checkpoints are available now on Hugging Face with no waitlist.