Krea 2 Technical Report
111 points
23 hours ago
| 5 comments
| krea.ai
| HN
pwython
46 seconds ago
[-]
Looking forward to playing with Krea 2, I use Z-Image Turbo daily -- it has replaced my stock photo subscriptions, for realism and illustrations.

May I ask how much did the training cost you?

reply
mattnewton
23 hours ago
[-]
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!

reply
ttul
2 hours ago
[-]
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.
reply
mattnewton
6 minutes ago
[-]
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.

reply
justinclift
3 hours ago
[-]
Interesting item on the careers page btw. For anyone that knows what older school Mellanox was about, it might be your kind of thing: https://jobs.ashbyhq.com/krea/ebe94024-eef6-4306-a019-10072a... :D
reply
kodablah
1 hour ago
[-]
reply
BoredPositron
41 minutes ago
[-]
It's a good model sadly the use of the qwen vae is a bit of a downer.
reply
mobiuscog
39 minutes ago
[-]
It's been mentioned by some that using the wan2.1 vae instead solves this. I haven't personally had time to try yet.
reply