Hi HN, we're releasing weights for our latest text to image model and publishing this writeup on how we trained it in quite a bit of depth.
I hope there is something in the report for everyone, we included a fair bit on the actual training and data infrastructure usually not written about much, that I think will be interesting to people here. There's more that didn't fit, happy to answer questions!
This is a massive technical report for an open weights image gen model. As someone who has followed this space closely, it’s really cool to read about the behind-the-scenes experimentation and effort that went into the final product. I hope you will release some of the find tuning tools so the community can experiment with them as well and really push what the model’s capable of.
You can find some links and details in the GitHub readme for finetuning / LoRA support. Ostiris, musubi tuner, fal and hugging face diffusers are all day-0 supported :)
https://github.com/krea-ai/krea-2
We recommend training off the undistilled, Raw checkpoint, and then applying the LoRA to the Turbo model for inference.
Neat! Between Ideogram4, Flux2, Qwen-Image, ZiT, and Krea - there's been a lot of positive movement in the open-weights space.
The original Flux.1 Krea is actually in my GenAI Showdown benchmark site from all the way back in July of last year (which feels like a lifetime in this space), so I’m looking forward to putting this new one through its paces.
What is Krea's approach to content such as pornography and gore? It's been frustrating to see all of the leading models take a very hard line on excluding vice content, even when it is perfectly legal, in the name of safety.
Krea 2 Large (on the website and api) was trained with the FLUX 2 VAE, if you want to test it out and push realism. After working with both I think the flux VAE has a slight edge in learning realistic textures but it's smaller than you might think, the Qwen VAE was overall very good in ablations and good at learning to produce a diverse set of styles.
We are releasing the weights and a _juicy_ technical report---at least given current industry standards. In it we describe data curation/captioning, model architecture, post-training, RL pipelines, prompt expansion, style references, and our infrastructure in great detail.
When it comes to theweights themselves, there's actually 2 releases:
* Krea 2 Turbo. This model is both guidance- and timestep- distilled for faster inference.
* Krea 2 RAW. This model is actually meant to be hackable/fine-tunable
One of the things we think the (open) LLM community does well is release models in different sizes and also at different stages of the training pipelines; we are releasing two checkpoints at both the mid-training and post-training stage. This is rare in the image & multimedia community, so we can't help it but to feel proud of this release.
Good to have more open weight models, and I really appreciate the in-depth write-up.
I also like the "keep the manifold wide" approach of trying to make a model capable of many styles as opposed to getting it "dialed in" for a dozen of style presets.
But it does feel very much like "fighting the past war" - now that advanced "image-to-image"/"agentic composition" models like Nano Banana 2 or Images 2.0 are out there in force.
I seriously doubt that the basic Qwen 3 VL in cross can get anywhere near that level of I2I. And robust I2I is very desirable - editing, adjustment, character consistency, the generalization of whatever you're doing with style transfer now (underexplained BTW).
Trying to hit that level of I2I is not by any means easy, but it's pretty clear to me that this is where the next frontier for image models lies. Feels like Ideogram might be building up to it, but I'm yet to see it anywhere else in open weight space.
Results are in! This is a really impressive showing especially given how fast the Turbo model at 8 steps. The only locally hostable model that managed to outperform it was Ideogram 4 which is significantly slower (think minutes vs seconds).
It did fall to the usual “model killers”: the nine-pointed star, Count Rugen, the overcrowded flat Earth. But overall, it really punched above its weight class, scoring the highest among locally hostable models and coming in just below Ideogram 4 passing 6 of the 15 tests.
Great job Krea team!
GenAI link to compare locally hostable models only:
Speaking of model killers, can they do “wine glass completely full to the top” yet? That‘s the one I used to show people but I haven’t tried it in a while.
In the launcher GUI, select 'Image Gen' and set Krea2 Turbo gguf as the image model, Qwen3 VL 4b gguf as the T5 model, and Wan2.1 safetensors as the VAE.
> 2.3 Revenue Threshold for Commercial Use. Commercial Use under this Agreement of the Krea Model, Derivatives, or Outputs is permitted 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), calculated on a trailing twelve-month basis and including all revenue from
all sources. If you meet or exceed this threshold, you must obtain a separate enterprise license from Krea prior to any Commercial Use. If your revenue meets or exceeds this threshold at any time during your use of the Krea Model under this Agreement, you must immediately cease Commercial Use and contact Krea. Enterprise license inquiries may be directed to opensource@krea.ai.
> 4.1 General Restrictions. You shall not, and shall not permit any third party to: (a) Use the Krea Model, any Derivative, or any Output in violation of applicable law, regulation, this Agreement, or the Acceptable Use Policy;
> 4.2 Content Filtering Requirement. You must implement reasonable and appropriate Content Filter measures to detect, prevent, and mitigate the generation or distribution of prohibited, harmful, or unlawful content through your deployment of the Krea Model or any Derivative. Such measures may include, but are not limited to: (a) open-source content classifiers, such as Falconsai/nsfw_image_detection, NudeNet, or CompVis safety checker; (b) commercial content
moderation APIs, such as Hive Moderation or Microsoft Azure AI Content Safety; (c) manual human review processes; and/or (d) any combination of the foregoing or other technically appropriate measures.
> 4.4 Acceptable Use Policy Compliance. You must comply with the Acceptable Use Policy, which is incorporated herein by reference.
> You shall not use or allow others to use the Krea 2 Raw Model or Krea 2 Turbo Model, any Derivative, or any Output for any of the following purposes:
> (8) Circumventing or removing any safety measures, usage restrictions, content filters, content provenance, or watermarking mechanisms implemented by Krea or any deployer;
22 comments
[ 6.7 ms ] story [ 67.9 ms ] threadI hope there is something in the report for everyone, we included a fair bit on the actual training and data infrastructure usually not written about much, that I think will be interesting to people here. There's more that didn't fit, happy to answer questions!
We recommend training off the undistilled, Raw checkpoint, and then applying the LoRA to the Turbo model for inference.
We also had 0-day support from people like Ostris and ComfyUI from the open source community
The original Flux.1 Krea is actually in my GenAI Showdown benchmark site from all the way back in July of last year (which feels like a lifetime in this space), so I’m looking forward to putting this new one through its paces.
May I ask how much did the training cost you?
I am Diego Rodriguez, Co-founder & CTO at Krea.
We are releasing the weights and a _juicy_ technical report---at least given current industry standards. In it we describe data curation/captioning, model architecture, post-training, RL pipelines, prompt expansion, style references, and our infrastructure in great detail.
When it comes to theweights themselves, there's actually 2 releases:
* Krea 2 Turbo. This model is both guidance- and timestep- distilled for faster inference.
* Krea 2 RAW. This model is actually meant to be hackable/fine-tunable
One of the things we think the (open) LLM community does well is release models in different sizes and also at different stages of the training pipelines; we are releasing two checkpoints at both the mid-training and post-training stage. This is rare in the image & multimedia community, so we can't help it but to feel proud of this release.
We are on par with Nano Banana in terms of image quality as per Artificial Analysis text-to-image benchmarks (https://artificialanalysis.ai/image/leaderboard/text-to-imag...).
We also attached a permissive license for individuals and small businesses.
Useful links:
- Marketing page around the OSS release: https://www.krea.ai/krea-2-open-source
- Huggingface model: https://www.krea.ai/krea-2/huggingface
- GitHub repository: https://www.krea.ai/krea-2/github
- Reddit AMA: https://www.reddit.com/r/StableDiffusion/comments/1udnm0a/we...
- Technical report: https://www.krea.ai/blog/krea-2-technical-report Thank you and I hope you enjoy this release---happy hacking!
Some of our team members will be answering questions since we are at the front page for now (thank you HN!).
Happy hacking!
I also like the "keep the manifold wide" approach of trying to make a model capable of many styles as opposed to getting it "dialed in" for a dozen of style presets.
But it does feel very much like "fighting the past war" - now that advanced "image-to-image"/"agentic composition" models like Nano Banana 2 or Images 2.0 are out there in force.
I seriously doubt that the basic Qwen 3 VL in cross can get anywhere near that level of I2I. And robust I2I is very desirable - editing, adjustment, character consistency, the generalization of whatever you're doing with style transfer now (underexplained BTW).
Trying to hit that level of I2I is not by any means easy, but it's pretty clear to me that this is where the next frontier for image models lies. Feels like Ideogram might be building up to it, but I'm yet to see it anywhere else in open weight space.
I tried two of the Krea 2 models in LM Studio, but loading the downloaded models errored out. (Maybe I'm doing it wrong, since it's an image model.)
Previously: https://news.ycombinator.com/item?id=47800562
It did fall to the usual “model killers”: the nine-pointed star, Count Rugen, the overcrowded flat Earth. But overall, it really punched above its weight class, scoring the highest among locally hostable models and coming in just below Ideogram 4 passing 6 of the 15 tests.
Great job Krea team!
GenAI link to compare locally hostable models only:
https://genai-showdown.specr.net/?models=fd,hd,kd,qi,f2d,zt,...
* https://github.com/LostRuins/koboldcpp
* https://github.com/LostRuins/koboldcpp/releases/tag/rolling
Model weights:
* https://github.com/leejet/stable-diffusion.cpp/blob/master/d...
In the launcher GUI, select 'Image Gen' and set Krea2 Turbo gguf as the image model, Qwen3 VL 4b gguf as the T5 model, and Wan2.1 safetensors as the VAE.
> 2.3 Revenue Threshold for Commercial Use. Commercial Use under this Agreement of the Krea Model, Derivatives, or Outputs is permitted 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), calculated on a trailing twelve-month basis and including all revenue from all sources. If you meet or exceed this threshold, you must obtain a separate enterprise license from Krea prior to any Commercial Use. If your revenue meets or exceeds this threshold at any time during your use of the Krea Model under this Agreement, you must immediately cease Commercial Use and contact Krea. Enterprise license inquiries may be directed to opensource@krea.ai.
> 4.1 General Restrictions. You shall not, and shall not permit any third party to: (a) Use the Krea Model, any Derivative, or any Output in violation of applicable law, regulation, this Agreement, or the Acceptable Use Policy;
> 4.2 Content Filtering Requirement. You must implement reasonable and appropriate Content Filter measures to detect, prevent, and mitigate the generation or distribution of prohibited, harmful, or unlawful content through your deployment of the Krea Model or any Derivative. Such measures may include, but are not limited to: (a) open-source content classifiers, such as Falconsai/nsfw_image_detection, NudeNet, or CompVis safety checker; (b) commercial content moderation APIs, such as Hive Moderation or Microsoft Azure AI Content Safety; (c) manual human review processes; and/or (d) any combination of the foregoing or other technically appropriate measures.
> 4.4 Acceptable Use Policy Compliance. You must comply with the Acceptable Use Policy, which is incorporated herein by reference.
The acceptable use policy is on the website (https://www.krea.ai/krea-2-use-policy) and includes:
> You shall not use or allow others to use the Krea 2 Raw Model or Krea 2 Turbo Model, any Derivative, or any Output for any of the following purposes:
> (8) Circumventing or removing any safety measures, usage restrictions, content filters, content provenance, or watermarking mechanisms implemented by Krea or any deployer;