Hacker News Reader

Comment by vunderba | Hacker News Reader

Nice release! The 12B open-weight Krea 2 scored pretty well in my GenAI benchmark [1] on txt2img (6 out of 15) when I tested it back in June, so I’m looking forward to putting Krea 3 through its paces.

While Krea 2 was principally a text-to-image model, it seems like Krea 3 might be flipping that around? When I tried the beta a few minutes ago, I noticed a warning message saying that text-to-image was experimental at this point. Is the plan still to treat text-to-image as a first-class citizen, or is Krea 3 going to be more focused on being an editing model?

[1] - https://genai-showdown.specr.net?models=local

View on HN

The honest answer is that the team is pretty happy with the text to image of Krea 2 and our team prioritized editing which was the big gap in the previous model’s capabilities.

We know users are used to using one model for everything though and have already turned our attention to the text to image side; that message is to let people know that the text to image performance is not final, we expect the final tuning of Krea 3 to beat Krea 2’s high bar there soon.

(for context, matthew worked first hand on the development of krea 3!)

> The honest answer is that the team is pretty happy with the text to image of Krea 2 and our team prioritized editing which was the big gap in the previous model’s capabilities.

I definitely agree. Part of my benchmark includes a "compliance score" which is effectively how easy it is to steer the model and Krea 2 ranks among the highest of the local models.

> We know users are used to using one model for everything though and have already turned our attention to the text to image side

Thanks for the answer. It definitely seems like open-weight unified models are becoming much more popular (Qwen-Image 2.1, Ming 0.1, upcoming Ideogram 4.5, etc).

Looking forward to it!

We will still treat t2i as first-class citizen too, but our initial work focused on edit because SO MANY people exclusively did editing, rather than “vanilla” generation.

But we’re opening the beta now (after polishing the initial jagged edges in a private beta rollout), to find even more edges now, but at least now the frontier-level capabilities and new capabilities should be available for people to get a great experience.

Do you have a workspace in Krea? I’m happy to put some API credits for you to test Krea 3.

I'd like to look into adding it to https://github.com/jcelerier/librediffusion for real time inference using tensorrt, do you think it's worth trying?

We’ve tried a lot of inference techniques but we’re not a real-time level yet (and believe us, we’ve done real time models!).

We’d like to open-source it soon (it’s just that it’s another can of worms - that we already went thru for Krea 2), precisely for people to try things like this. The Krea 2 community efforts have been incredible to watch from the inside.

Once Krea 3 is open, you should def try!

Hi dvrp - Appreciate the answer definitely glad to hear it. I split my time pretty evenly between text-to-image and image-to-image, so seeing another unified model is exciting!

I didn't (but I do now) have an account on Krea since I run most of my benchmarks through a bespoke suite of tools via APIs like OpenRouter and Fal.AI, but I appreciate the generosity! I'll reach out to you via your profile~

Reply on news.ycombinator.com