dvrp

I think for everyone interested in this project checking your OG twitter thread is also somewhat mandatory:

https://x.com/_chenglou/status/2110557373651931267

It took me a bit to wrap my head around what you did.

Ladies and gentlemen, the Streisand Effect in its full glory!

Did you check the “Only what you asked. Not one pixel more” section?

Not only do I think the website section addresses exactly this, but we also built a pixel diff viz tool for you to compare with other models.

I meant workflows in the generic definition. Not as in Node-based Workflows (which we also support).

EDIT: Our engineering staff realized you can play games with Krea 3. Or like they like to call it “the most expensive game engine in the world”

See: https://x.com/asciidiego/status/2110846676139221376?s=20

they’ll come!

We will release a technical report about it. But our Krea 2 technical report is out and, AFAIK, it’s the most open technical report out there (1 hour estimated of reading time lol) so you can get a lot of information there already!

Thank you for the constructive reply; I think you are quite right.

I’ll ponder about what would be a better growth strategy - even in turbulent times - while, like you say, we also ensure is done on a good financial base/cushion.

Thanks for taking the time too.

correct!

I think my product team is rapidly iterating on this page based off feedback and the HN crawler it’s re-scraping metadata such as title tags

the post used to be titled “Frontier Image Model 4x faster than Nano Banana Pro”

I actually was just talking about this with a peer. He was saying that people inside labs and data brokers use image models for RL training of LLMs to get better at these sort of tasks, which would make sense as a way to try to embed the knowledge into the LLM - similar to how Astra was clearly trained with more CAD, 3D, and Robotics in mind.

Oh! We should put it there then!

I was aware of AWS Bedrock

Check Krea 2. Not only open weight but even better terms in the license!

Also, Krea 1 was co-developed with BFL !

K3 open weights is being discussed internally as I type this!

Gemini (the product) is a mix of an LLM plus image model plus agentic tooling. Is not apples to apples.

A fairer comparison would be Krea Agent with Krea 3 as image model. So you could have grounding, web search, verification models that assess results, etc.

If you told me Nano Banana Pro and that’s it, that’s a different story.

Ngl that made me laugh

Yes!

And regarding your use-case: we want to let people do LoRAs within the platform too right away like we did with Krea 2 -- so it should be possible right now if you want to try, both locally or in the app!

Businesses need more than just cost per image than a raw ChatGPT subscription doesn’t give you.

There’s also workflows around non-frontier models that make sense, especially at-scale (think internationalization of a marketing campaign to 200 countries across different media formats). A lot of the same rules that apply to LLMs apply to diffusion. The same argument could be made around Codex/Astra/Sol and many people are happily paying for tools like Devin/Cognition.

But I agree with you that if you are not careful, the subsidies from big labs can eat your business.

EDIT: I hope you’re enjoying Krea 2! What GPU are you running it on?

As in that they don’t look like fashion models?

Exactly, it seems to me like a workflow that is not “Bitter Lesson”-pilled and was a workaround non-capable LLMs.

We are also seeing it with people using Opus 5.5 to code shaders, motion graphics, and websites more as opposed to using, say, after effects.

Thank you haha

And benchmarks do not account for it sadly

Also, sometimes you want to do “ugly” things on purpose or add imperfections like film grain and such

oh yea, we are working on enabling people to fine-tune it

yes on gpt-image 2.5 being great at times, at almost 1 order of magnitude more of the cost per image and way slower tho

EDIT: do you find that a lot of people use diffusion models for UI mockups? i’ve always been surprised that people wouldn’t use LLMs exclusively for that instead

We focused partly on the frontier ones that most people use or recognize. We have some usage data from free users as well as what people talk in social media.

But there’s a whole plethora of worthy models (e.g. flux family, seedream, qwen, etc.) that are not mentioning purely cuz marketing and to avoid complicating the site too much. We hope to release a more technical report soon.

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!

Maybe silly question but what’s model garden?

Yea, that’s what I used to think until I became a business owner in the generative AI world, and until other competitors and people in our space start doing sneaky practices that, verifiably, hurt your business and you’re forced to defend yourself. Including sending bot farms to your personal Instagram accounts and even those of your family and friends.

This mostly doesn’t happen in Business or Enterprise tiers, but in consumer is ruthlessly competitive.

Don’t hate the player and so on. We don’t like it particularly either.

EDIT: I maybe should probably put an aggressive coupon code, but even through Hacker News I’ve gotten email-spammed because of LLMs that do sentiment analysis and similar techniques. Such a tragedy of the commons if you ask me.

oops. great catch. fixing rn!

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

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.

Hello HN,

This is Diego from Krea. Today we are announcing the open beta of Krea 3, our newest foundational image model.

In a nutshell, it’s a frontier image model in terms of quality, but at a fraction of the speed and a fraction of the price.

We will be releasing a more technical update, like we did with Krea 2 (), but this website shows the capabilities we focused on:

* Native regional edits

* Precise edits so that there is minimal area degradation pixel-wise

* Frontier image quality. Benchmarks still being worked on, but in our internal evals it seems a no-brainer replacement to NBP or GPT Image, both speed and cost-wise.

* Faster edits (AFAICK, this should unlock a pareto frontier for speed in frontier edit models).

In terms of future updates, the team is thinking about letting folks fine-tune Krea 3 (via LoRAs) and open-weights, but we’re still in the middle of launch hiccups and then we’ll revisit the next iterations.

Thanks!