Skip to content

Note krea-2

All posts
Nov 17, 20269 min readkrea-2 · text-to-image · open-weights · ai-image-generation · open-source · lora · comfyuiRead in:ENID

Krea 2: The Open-Weights Text-to-Image Model You Can Run Locally at 2K

Krea 2 launched in June 2026 with two checkpoints, RAW for training LoRAs and Turbo for fast 2K generation. It has 982 stars on GitHub, and its license lets small teams use it commercially for free.

A glowing red futuristic processor cube on a black grid floor. Photo by Pachon in Motion on Pexels.

Most image models ask you to choose: pay for an API, or settle for a smaller open model that cannot keep up. Krea 2 tries to remove that trade-off. It is an open-weights text-to-image model from Krea AI that ships as two checkpoints, RAW for training and Turbo for fast generation, and the Turbo checkpoint generates images at up to 2K resolution in only 8 steps. The official inference code lives on GitHub as krea-ai/krea-2, where it has 982 stars. (GitHub)

This post explains what Krea 2 is, how the two checkpoints differ, what the license actually allows, how it scores on the Artificial Analysis leaderboard today, and what hardware you need to run it on your own machine.

Key Takeaways

Krea 2 is an open-weights text-to-image model from Krea AI, released June 22, 2026, with 982 stars on GitHub as of November 17, 2026.

It ships two checkpoints: RAW for training LoRAs (52 steps) and Turbo, an 8-step distilled model that generates images at up to 2048x2048.

The inference code is Apache-2.0, but the model weights use the Krea 2 Community License: commercial use is free only under $1 million in annual company revenue.

At launch it ranked 2nd among independent labs on Artificial Analysis; as of November 17, 2026, Krea 2 Large sits at rank 31 with an Elo of 1035.

It is a 24GB-card model in practice, but a 16GB card can work at 1 megapixel with the text encoder offloaded, according to independent estimates.

What Is Krea 2?

Krea 2 is a text-to-image model built on a 12B dense DiT backbone, a Qwen Image VAE, and a Qwen3-VL text encoder with multi-layer feature aggregation. Krea announced the model on May 12, 2026, and released the open weights on June 22, 2026. (Krea AI)

The architecture is deliberately practical. The Qwen3-VL text encoder understands detailed prompts and image references, the diffusion transformer does the image synthesis, and the VAE turns latent representations into pixels. Krea trained it in stages: pretraining at 256px, 512px, then 1024px on billions of images, followed by midtraining, supervised fine-tuning, preference optimization, and reinforcement learning. One detail worth noting: Krea says it used no AI-generated images in its pretraining mix. (Krea AI)

Watch the 30-second repo signal on Repomaniac to see it in action.

Two Models in One Release: RAW and Turbo

Krea 2 is not one checkpoint, it is a workflow. RAW is the undistilled base model, built for training and research. Turbo is an 8-step distilled version of the same model, built for fast generation. Krea's own slogan for the pair is simple: "Train LoRAs on RAW, then run them fast on Turbo." (Krea AI)

Aspect

RAW

Turbo

Purpose

Train LoRAs, research, fine-tuning

Fast image generation

Inference steps

52

8

Guidance (CFG)

3.5

0.0

Max resolution

Up to 1K

Up to 2K (2048x2048)

Best for

Custom styles and characters

Everyday generation

The split matters because most open models force a compromise: a base model that is slow, or a distilled model you cannot train on. Here you get both from the same weights family, and a LoRA trained on RAW runs on Turbo without retraining. Recommended fine-tuning tools include Hugging Face Diffusers, Ostris AI Toolkit, Fal, and Kohya musubi-tuner, and the model runs in ComfyUI, Fal, and SGLang. (GitHub)

The License Split You Need to Know

This is the part most summaries get wrong, so read it carefully. Krea 2 has two licenses, not one.

The inference code in the GitHub repo is Apache-2.0, a permissive open source license. The model weights, however, are under the Krea 2 Community License v1. Commercial use of the weights is allowed only if your total company-wide annual revenue is under $1,000,000 USD over the trailing 12 months, counting all affiliates. Above that threshold you need an enterprise license from Krea (opensource@krea.ai). The license also requires content filtering, and the grant can be revoked on 30 days notice. (GitHub)

In practice, that means hobbyists, freelancers, and small studios can use Krea 2 commercially for free, while larger companies need to talk to Krea first. Do not describe Krea 2 as simply "Apache-2.0 licensed": that is true of the code, not of the weights.

How Good Is It Really?

Krea 2 launched strong. In June 2026 it placed in the top 10 overall on the Artificial Analysis image leaderboard and 2nd among models from independent labs. That launch result is still quoted in the README today, but the leaderboard has moved on, and the current numbers tell a more honest story. (Krea AI)

Model

Rank (Nov 17, 2026)

Elo

Price per 1k images

GPT Image 2.6 Sunburst (max)

1

1261

Not listed

Qwen-Image-2.1

28

1047

Not listed

FLUX 3 Image Klein

29

1041

Not listed

Krea 2 Large

31

1035

$60.00

Krea 2 Medium Turbo

45

1020

$15.00

Krea 2 Medium

50

1014

$30.00

FLUX.2 [dev]

54

1000

Not listed

As of November 17, 2026, Krea 2 Large sits at rank 31 with an Elo of 1035, based on 15,164 samples. That puts it just behind Qwen-Image-2.1 and FLUX 3 Image Klein, and well behind the closed leader, GPT Image 2.6 Sunburst at 1261 Elo. The ranking is a snapshot, not a verdict: Krea 2's appeal was never only about Elo, it is about open weights, local inference, and a LoRA ecosystem that closed models cannot offer. (Artificial Analysis)

Running It Locally: What Hardware Do You Need?

Krea publishes no official VRAM requirement, so treat every hardware figure below as an independent estimate, not a manufacturer spec. The numbers come from BestFreeWebResources, which measured file sizes and calculated VRAM from them. (BestFreeWebResources)

Component

File size

Model weights, bf16

26.28 GB

Model weights, fp8

13.14 GB

Model weights, NVFP4

7.67 GB

Qwen3-VL 4B text encoder

5.24 GB

VAE

0.25 GB

Standard template set

About 18.6 GB

Estimated VRAM at 1 megapixel: 38.2 GB in bf16 with everything in VRAM, 20.6 GB in 8-bit, 16.2 GB in 8-bit with the text encoder offloaded, and 10.7 GB in 4-bit with sequential offload. The practical verdict: Krea 2 is a 24GB-card model, and a 16GB card can work at 1 megapixel once ComfyUI moves the text encoder off the GPU. If you want 2K output, expect to need the larger card or accept slower offloaded generation.

LoRAs and the Style Ecosystem

Krea 2's fastest growing asset is its LoRA scene. Krea ships 9 official style LoRAs, each with its own trigger word at strength 1.0: Darkbrush, Dotmatrix, Kidsdrawing, Neondrip, Rainywindow, Retroanime, Softwatercolor, Sunsetblur, and Vintagetarot. (ComfyUI Wiki)

The community has gone further. As of November 2, 2026, there are 933 Krea 2 LoRAs on Hugging Face, which one comparison calls "the fastest-growing LoRA scene of 2026." For context, Qwen-Image 2.1 has 7 LoRAs and FLUX 3 Image Klein has 31, though the long-established SDXL still leads by far with about 9,800. (FreakyJolly)

That ecosystem is the real reason to pick Krea 2 over a slightly higher ranked model. A LoRA lets you lock in a character, a product, or a house style, then generate consistent images in seconds on Turbo. For illustrators and small teams, that consistency is worth more than a few Elo points.

Krea 2 vs Other Open Image Models

Aspect

Krea 2

Qwen-Image 2.1

FLUX 3 Image Klein

SDXL 1.0

Parameters

12B class

7B generator

5.6B

3.5B

License

Community ($1M revenue cap)

Qwen Research (non-commercial)

Under $5M revenue

Open RAIL

VRAM, practical

12-16 GB (fp8/nvfp4)

16 GB

8 GB

8 GB

LoRAs on Hugging Face

933

7

31

About 9,800

Best for

Styles, LoRAs, 2K local generation

Detail, editing

Lightweight local use

Mature ecosystem

The table shows the trade-off clearly. Krea 2 is the heaviest model here, but it is the only one with a genuinely commercial-friendly license for small teams and a large, growing LoRA library. Qwen-Image 2.1 scores slightly higher on the leaderboard but its research license blocks commercial use. FLUX 3 Image Klein is the lightest to run. SDXL remains the ecosystem king, but it is a 2023 model, and it shows in image quality.

If you want another open image model to compare, read our spotlight on Ideogram 4, the open-weight model that finally gets typography right.

How to Get Started

The quickest path is ComfyUI, which has an official Krea 2 package. Install ComfyUI, add the Krea 2 nodes, download the Turbo checkpoint in fp8 or NVFP4 plus the Qwen3-VL text encoder and the VAE, then load a workflow and generate. For training, start from the RAW checkpoint with Diffusers or the Ostris AI Toolkit, train your LoRA, then run it on Turbo for fast inference.

Two practical tips. First, start with Turbo in fp8: it is the fastest way to see what the model can do, and you can switch to RAW later for training. Second, check the license before you ship anything commercial: if your company's annual revenue is over $1 million, contact Krea about an enterprise license first.

Frequently Asked Questions

Is Krea 2 really free?

The code is Apache-2.0 and the weights are free for commercial use under $1 million in annual company revenue. Above that you need an enterprise license. The model itself costs nothing to download and run locally.

Can I run Krea 2 on a consumer GPU?

Yes, with caveats. Estimates put it at 12 to 16 GB of VRAM in fp8 or NVFP4 at 1 megapixel, so a 16GB card works with the text encoder offloaded, and a 24GB card handles it comfortably. For 2K output, a 24GB card is the safer choice.

What is the difference between RAW and Turbo?

RAW is the undistilled base model: 52 steps, CFG 3.5, built for training LoRAs and research. Turbo is an 8-step distilled checkpoint: CFG 0.0, built for fast generation at up to 2K resolution. LoRAs trained on RAW run on Turbo.

How does Krea 2 compare to FLUX and Qwen-Image?

On the Artificial Analysis leaderboard as of November 17, 2026, Krea 2 Large (Elo 1035) trails Qwen-Image-2.1 (1047) and FLUX 3 Image Klein (1041) by a small margin. Its advantages are the commercial-friendly license for small teams, the large LoRA ecosystem, and 2K local generation.

Can I train my own style with Krea 2?

Yes. Train a LoRA on the RAW checkpoint with Diffusers, Ostris AI Toolkit, Fal, or Kohya musubi-tuner, then run it on Turbo for fast inference. Krea ships 9 official style LoRAs to start from.

Conclusion

Krea 2 is not the top-ranked image model anymore, and it does not pretend to be. What it offers is rarer: genuinely open weights, a two-checkpoint workflow that separates training from fast generation, a license that lets small teams use it commercially for free, and a LoRA ecosystem that grew to 933 community styles in under five months. If you generate images locally and want a model you can shape to your own style, Krea 2 is one of the strongest open options available today.

Download it from GitHub, try the Turbo checkpoint first, and train a LoRA on RAW when you find a style worth keeping.

Baca versi Bahasa Indonesia dari artikel ini di Krea 2: Model Gambar AI Open-Weights yang Bisa Dijalankan Lokal di Resolusi 2K.

Tags: krea-2 · text-to-image · open-weights · ai-image-generation · open-source · lora · comfyui

All posts