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.
It's pretty great that you are providing the undistilled model on day 0. Here's a pro-tip: With Flux.2 Klein, someone created a turbo slider LoRA - basically a diff of the turbo 9B model vs. the undistilled 9B model. What's great about this LoRA is that you can sample using a heavier weighting of the undistilled weights during early sampling steps and then finish the sampling off with mostly the distilled weights. The result is a better "finish" (taking advantage of the distilled model's refinement for image quality) without sacrificing the undistilled model's greater ability to adhere to the prompt, because the undistilled model doesn't have to devote its weights so much to looking good.
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.
> What is Krea's approach to content such as pornography and gore?
You mean whether it's censored? Of course it is. They even say so themselves[1]:
> The open source version needed to go through some alignment training so there might be some inconsistencies between closed / open version.
The "alignment training" is pretty much a code word for censorship/lobotomy, because unlike the API version they can't slap on a safe/unsafe classifier on an uncensored model. Of course, just to be clear: I don't blame. If they don't do this then the next post we'd see is them being in hot water. (Remember what happened with Grok? That's what happens when you don't take a hard line here.)
> If they don't do this then the next post we'd see is them being in hot water. (Remember what happened with Grok? That's what happens when you don't take a hard line here.)
I disagree.
Grok is a hosted service that can use classifiers and has been personally guided by Elon Musk directly before, that was choosing to generate and distribute child porn for users. Then after getting in trouble, had the gall to continue doing it behind a paywall.
that was choosing to generate and distribute child porn for users.
This is a slant, a lie of presumption and ignorance.
Ignorance of what "open" means, presumption of assigning intent without validation.
Some believe that lobotomizing models leads to suboptimal results overall. Some also believe that the user, not the tool is responsible.
In these contexts, you'd better call a paint manufacturing company "choosing to enable child porn".
I am so sick of US style politics, with its insane labelling by both parties, and their worshippers, with intent to besmirch on every breath, because someone doesn't agree with them in some narrow political context.
From outside the US, most can barely tell the difference between a democrat and a republican. Both sides are slimy, smarmy, dishonest, without honour, and spew lies and bull about anyone they don't agree with.
If one commissions an artist to paint child porn, the artist is still the one guilty of producing and distributing child porn. If I contract a company to host a child porn website, they are still hosting a child porn website!
Furthermore, even if I don't make the painting, if I see it and distribute it anyway, I'm still distributing child porn. You don't get to just say "whoopsies, I didn't make it, how could I possibly know the open and visible image was illegal?"
Grok isn't making paint. It's not making monitors. It doesn't send you graphics software. It itself produced pornographic images of real children and distributed them to pedophiles for profit.
And even if we wanted to give them the benefit of the doubt, after this came to light and they got in trouble, they disabled this capability for free users while allowing paid users to continue. They knew what they were doing at that point, so no more ignorance plea.
At least open model producers are just distributing software that can produce illicit images. But because they don't actually do it, they can wash their hands of responsibility.
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.
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!