Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Krea 2: SOTA open-weights 12B image model

krea.ai

Post reply on HN
Print view View original

21–30 of 50 posts

Re: Krea 2: SOTA open-weights 12B image model

#21

Post by refulgentis »

ACCount37 wrote: post #20

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…

This model does image to image; whats the issue with Qwen 3 VL; is style transfer unexplained? " reference" is mentioned 11 times on the page (more specifically, I read it and it seemed to discuss it a lot)

Re: Krea 2: SOTA open-weights 12B image model

#22

Post by dvrp »

ACCount37 wrote: post #20

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…

I appreciate the skepticism but we find internally that this model is used more than Nano Banana for many cases like moodboarding (also, 4x cheaper than NBP never hurts). Agentic workflows are compatible with Krea 2 so I’m not sure I follow there. If you are talking about an edit model, that’s coming too.

Also, we are on par with them in t2i benchmarks, check the artificial analysis link I posted in my top comment.

And you cannot re-train nano banana or ChatGPT to understand your brand, which is what our customers complain about constantly.

Plus open-source! It’s hard to do an apple to apple comparison.

Re: Krea 2: SOTA open-weights 12B image model

#23

Post by vunderba »

mattnewton wrote: post #2

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!

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.

Re: Krea 2: SOTA open-weights 12B image model

#24

Post by ttul »

mattnewton wrote: post #8
ttul wrote: post #4

Earlier quoted context omitted.

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.

Re: Krea 2: SOTA open-weights 12B image model

#26

Post by Taek »

mattnewton wrote: post #2

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!

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.

Re: Krea 2: SOTA open-weights 12B image model

#27

Post by ACCount37 »

dvrp wrote: post #22
ACCount37 wrote: post #20

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…

I appreciate the skepticism but we find internally that this model is used more than Nano Banana for many cases like moodboarding (also, 4x cheaper than NBP never hurts). Agentic workflows are compatible with Krea 2 so I’m not sure I follow there. If you are talking about an edit model, that’s coming too. Also, we are on par with them in t2i benchmarks, check the artificial analysis link I posted in my top comment. A…

"Compatible" is one thing - "built for" is a different beast. The difference can be like that between Images 1.0 and Images 2.0 - the sheer leap in compositional capabilities was staggering.

"Edit model" is a part of it, yes. So is style transfer. But less as an endpoint and more of a subset of what advanced I2I enables.

"Re-train to understand your brand" is a fine marketing pitch, but in practical terms, it's hard to justify burning a LoRA for most uses. Enthusiasts absolutely do it, but enthusiasts are built different. Robust I2I can accomplish a lot of the same, but with a workflow that's closer to "drag and drop your references" than to "try to get a LoRA to do what you wanted it to do on a very slim set of images".

Modern LoRA pipelines are getting closer to "reliable" and "braindead simple", but you can't escape the "wait N hours for the GPUs to churn" of fine tune no matter what you do. And iteration time kills - a lot of the value of AI in workflows is that it does what it does fast and allows you to iterate at speed.

You can think of "LoRA vs I2I" as of an image twin of "SFT vs in-context learning" of LLM land. Both are useful, neither substitutes for the other fully, but there's a reason why most reach for the latter way before they reach for the former.

I like the T2I from what I've seen, mind. Perhaps more than Images 2.0 or even NB2. I just think that focusing solely on T2I to the exclusion of advanced editing and composition capabilities is a very 2024 thing.

Re: Krea 2: SOTA open-weights 12B image model

#29

Post by Catloafdev »

realityfactchex wrote: post #28

Did anyone generate a pelican riding a bicycle? 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

Image models need software like ComfyUI, not LM Studio.

Re: Krea 2: SOTA open-weights 12B image model

#30

Post by dvrp »

realityfactchex wrote: post #28

Did anyone generate a pelican riding a bicycle? 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

here u go: https://ibb.co/XfZSJKyz

you can try it right away at krea.ai/image (warning: you need to sign-up)

Post reply on HN