* 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.
Does this mean "thinking about" open weights in the sense of remaining work to do before releasing them, or is it whether to release them at all? hopefully it's the former, Krea2 was awesome because it was open
Just a heads up, the trial period the email you send out for new signups seems to be broken - which is making trying this out hard.
your email:
>I just unlocked a personal discount of 40% on your account, and switched Pro on for the next 3 hours so you can use it before you decide. It unlocks you all premium features as well as:
∞ No credit ceiling
But after using doing one round of three images with the beta, it won't let me generate another example. Trying to put a handful of tests through to compare to what we get from our current system on top of nano bannana pro.
As a Krea2 local model enjoyer, your challenge is that on the $20 GPT plan you get a fairly good amount of Codex use and image gen. If I used Krea, it would be because I fine tuned/built a workflow on top of it to do something GPT image couldn't/wouldn't do. Maybe there's a market for the cheapest "good enough" bulk image gen model (spam? idk what the applications would be), but personally I'd say you should focus on workflow add-ons/one click fine tuning/etc that design companies can use to create assets with their custom "look" since the GPT-Image look will become blase from overuse.
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?
I think the early models distorted the perception of what artists actually need. It's not about workflows. Image creation is inherently iterative. None of the platforms seem to understand that, while AI artists have somehow fallen in love with the unnecessary complexity of node-based tools. Apparently, 400 tangled noodles look impressive when the boss peers over your shoulder.
The big studios and agencies are all building custom tooling as they always have but now a lot faster.
I've seen a graphical designer struggle to use an analogous product. They tried prompts like "remove background from this photo of dancers, don't change their face": the model couldn't do it: it always changed the faces at least a bit. In other words, the model was not removing the background, instead it was generating the whole picture anew, but without the background, which not only seemed inefficient, but was also way too lossy.
Why do all the "frontier" models still occasionally output images with the horrible ai texture tells? I'm surprised whoever made the front page let those ones in lol. My hypothesis is that it is either an unpatching or feature loss shortcut/artifact. Until then I have no hopes of ai generated images being viable for anything actually beautiful or creative
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?
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.
> 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).
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.
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~
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.
> but at a fraction of the speed and a fraction of the price
"at a fraction of the speed" is generally understood to mean that it is much slower, but I don't think that is what you are intending to say (considering the title and website claim it is multiple times faster)?
Overall Krea 3 isn't near GPT-Image-2.5 in terms of visual creativity in this space, but it still should be commended for its ability to render small text details. That said despite it specifying you can add exact hex codes as color reference, I've found that they tend to leak into the design (check the valuation panel of the banking website - notice to the right of the "fixed indexed 7-yr" there's a hex code that made it into the image).
I'd love to get my hands on this to post-train and see if it can be pushed further in this space. I actually see a lot of potential for the model, despite the clear differences in the examples above.
That Banking website is a good example of GPT-Image-2.5 having absolutely no self control, to a point I am just shocked they havent been working to fix it. It's like the first thing the model does is "generate 100 slogans" and then try and figure out how to cram them all into the image.
You see it constantly, and its worse with GPT than Gemini models, but they both do it. They seem to think they need to tell not show. They cant not annotate their thoughts into the image. A lot of the time you see either random signs on posts added, tshirts with text describing the story, or hand written annotation over the image.
There's really no reason this should be a problem anymore on default settings, or that everybody should still be getting the exact same style of edgy poster out of GPT with icons for bullets, and lists that add nothing jammed into the image. They need to have some kind of seed library of templates they start from, and higher temperature on selecting different styles.
A reasonable conclusion, but you don't have all the data here in this particular instance. These are all part of a set of benchmark prompts I have that are designed to try and push the images to their extremes. The banking one is dense because the prompt itself specifies each of these things - the goal with this test is to see how well it can handle many different UI elements. Here's an excerpt from the json blob that is used as the prompt:
"subjects": [
{
"description": "A transparent header spans the top of the teal hero. The white ANNUI wordmark sits at the upper left, followed by Product, For Agents, Resources, and Company links. For Agents is underlined as the active page. Sign in appears near the upper right beside a large white rounded Get started button.",
"position": "top edge, full width, approximately 8% of viewport height",
"subject": "Navigation Header"
},
{
"description": "Left-aligned hero content includes the spaced eyebrow 'FOR AGENTS', a large cream serif headline reading 'Your clients’ brighter tomorrow.', supporting copy about modern financial software, real support, and helping independent agents grow their annuity business, plus a white pill-shaped 'Build with Annui' button and a circular play control labeled 'Watch for Agents (1:30)'.",
"position": "left half, upper-middle through lower-middle, approximately 46% width and 58% viewport height",
"subject": "Hero Copy and Actions"
},
{
"description": "A monumental classical female statue dominates the center-right of the hero. The figure has a pale blue-green face, dark teal hair, a radiant gold helmet and laurel details, and layered gold and blue garments. Architectural columns and cloudy sky fill the background, creating a sculptural financial-services editorial scene.",
"position": "center-right, extending from the top edge to the bottom of the hero, approximately 55% width and 85% viewport height",
"subject": "Classical Statue Hero Artwork"
},
{
"description": "A narrow text panel overlays the far-right architectural columns with vertically spaced uppercase text reading 'INDEPENDENT ADVICE A BRIGHTER TOMORROW', a short horizontal divider, and 'BUILT TOGETHER'.",
"position": "right edge of hero, upper-middle through lower-middle, approximately 15% width and 42% viewport height",
"subject": "Vertical Message Panel"
},
{
"description": "A full-width warm off-white strip beneath the hero contains four evenly spaced benefit columns separated by thin vertical rules. Each column has a pale beige circular icon and dark text: 'Streamline Sales' with quote and compare copy, 'Serve Clients Better' with carrier and education copy, 'Grow Your Practice' with tools and marketing support copy, and 'Stay Independent' with client, advice, technology, and support copy.",
"position": "below the hero, full width, approximately 18% viewport height",
"subject": "Agent Benefits Strip"
}
],
That's not to say that GPT-Image-2.5 doesn't have a certain smell to it though. It definitely has its isms.
I very firmly disagree. IMO image models are extremely under-used in this area.
Just for context, I ran design systems at Figma for 5 years / led their first prompt->design stream before quitting to make my own startup in this space.
Image models are trained on visual creativity, color usage, whitespace, etc - all the ingredients you need for good design principles. Where they struggle is often with information density and text rendering, which LLMs excel at. You can use the best parts of both of them though - use an image model for the initial design and an LLM to convert it to code.
One thing that's very telling is with openai's sites feature, in its chain of thought it actually renders out an image first as a design reference, then builds towards that.
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.
Or the gpt-image "fried" texture. Interestingly the best approach I've found so far is to use gpt-image for the initial design, but nano banana for the asset export once I have the design. IE in the case of the "Annui" design, I'd take that design and its backing json structure, hand it off to Opus to build it in html, then use nano banana to build out the images of the statues/background images.
You just gotta use the best parts of all of them :)
Which is interesting... after gpt-image-2 came out with this bug replacing the piss yellow filter, it seems like every other release except Google's have the same issue.
Off the top of my head, this exact artifact appeared in Qwen, Grok, and now Krea.
It makes me wonder if gpt-image-2 is being used to bootstrap post-training, or if their images are just so ubiquitous it's turned into a sort of "virus".
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
> 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
Not yet, but it's what I quit my role at Figma to make a bet on for my startup. In my experience diffusion is a far, far better creative designer than LLMs. LLMs are trained for consistent output, which makes sense for a billing function, but leads to some very obvious smell when doing frontend work (take claude design as an example - it's often very obvious what comes out of it). Image models have a bit more creativity baked in, and are trained on color usage, whitespace, and visual balance a lot more. I've found the image->html approach leads to much better results, basically using an image model as an architect.
Would love to be connected with the team for when you all are ready to finetune it. My email is j@diffui.ai
When VL models started becoming popular a while back, a relatively common workflow was to use generative models (like the original Nano Banana) to generate initial UI mockups and then throw the images at a VL model to convert them into CSS and HTML.
However feels like LLMs such as Opus 5.5 in particular are surprisingly adept at generating UI mockups right off the bat these days. And since they’re already code, they’re much easier to modify and tweak.
I feel the same. For small motion graphics I am using claude in code and I think I didnt open AE for weeks. Bt when you need to put real image do scene (for example background or character) and it has to look same in 5-10 shots. Can krea keep it in same style/position through all edits? Or it can drift a bit?
Thank you. This shows that Krea 3 website isn't being honest, and this is what I hate about all those AI tools: the cherrypicking. Stop wasting our time!
And honestly, the whole thing looks like an ad-for-an-ad.
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!
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!
In the USA, if you "subscribe" to something, that is a promise to pay on the regular, until you cancel or unsubscribe. It's not a "single purchase".
If you cancel your card, or use a virtual one that's immediately invalidated, the service will have every right to come after you, and bill you for subsequent installments. You will run up a bill. You can choose not to pay it, but you cannot say that you don't owe it, because you agreed to pay the bills in perpetuity until cancellation.
I have no idea why people claim this, and think that playing a shell game with their finances will stop it.
I’m sorry, but you are the one spreading misinformation.
But I’ll give you an easy way to prove that what you said is true instead of just yelling at each other.
Give me a single example in the US of an online service company providing service after a credit card declined on a no-commitment service with auto renewal that actually sued in court and recovered something. There is none.
Sneaky pricing. Offers you plans when you sign up, if you decline them, it offers you a discount.
Just offer the discounted price up front. If your business cannot survive without sneaky pricing, it's not a good business.
I detest this kind of pricing. It seems that for every subscription I need to threaten to leave to get a discount, and now need an AI to automate that process.
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.
> 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.
As a serial tech entrepreneur with a few decades of hands-on experience both in my own startups and also spending some years as a VP / product group owner inside an F500 acquirer, you may want to think more deeply on this. A hidden downside of modern A/B testing analytics is how easy they make it to min/max short-term gains while masking delayed and second-order effects.
In this instance, no doubt the analytics show this bumps ASP in the short-term but those analytics can't show you the downstream impacts. These tend to accrue over time creating an overall drag on the business that you have no way to identify. A good case in point is how Adobe successfully min/maxxed their way to a large portion of their customers either not trusting (or hating) them. The brand damage is now likely irreparable. Do what you must to make payroll but you're on a slippery slope. Be careful you don't let 'any dark pattern is okay, as long as it's legal and works' become status quo. The consequences can accrue faster than you can exit.
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.
The problems you desribe are real and you should talk more about them in general, but 'don't hate the player hate the game' always comes off as a flippant dismissal. You have a good product, I think you can do better than that.
That's a bullshit excuse for this particular issue. No one is forcing you to employ this specific dark pattern regardless of how you claim your competition behaves.
Gave it a photo of a woman wearing a T-shirt, asked it to seat her at a fancy bar. It invented a suit jacket, which makes sense, but the drawing on the shirt runs right over the jacket. Gemini did much better.
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.
I understand what you're saying. My first thought is "is that really applicable to a mistake like the text on the T-shirt extending over the suit jacket that partially covers it?" Maybe it is. I also don't know what you actually get when you prompt Krea 3 via their website as I did today. I don't know whether that folds in some guidance from an LLM first. Tough to say.
Why is there no comparison with FLUX 3 Image? I am not a big user of image models so maybe those are in different categories (or maybe it's not a relevant frontier model?). Just curious.
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.
"cre-" is a latin root meaning (more or less) "make" which is where we get "create" - both are likely drawing inspiration from that (as would be the case for Creo, a common CAD software).
its the service in Google Cloud for accessing third party AI models hosted in Google’s data centers, alongside Google’s first party models
it also has enterprise level considerations that aren't present in direct API access of Google’s models, rendering almost all of the go to API documentation irrelevant, so it just comes from experience to know to use model garden’s implementation
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 (https://www.krea.ai/blog/krea-2-technical-report), 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!
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
Does this mean "thinking about" open weights in the sense of remaining work to do before releasing them, or is it whether to release them at all? hopefully it's the former, Krea2 was awesome because it was open
Krea2 was released because the open model was an interims model. The hosted krea2 used the flux2 vae and open weights qwen1.
Just a heads up, the trial period the email you send out for new signups seems to be broken - which is making trying this out hard.
your email:
>I just unlocked a personal discount of 40% on your account, and switched Pro on for the next 3 hours so you can use it before you decide. It unlocks you all premium features as well as: ∞ No credit ceiling
But after using doing one round of three images with the beta, it won't let me generate another example. Trying to put a handful of tests through to compare to what we get from our current system on top of nano bannana pro.
As a Krea2 local model enjoyer, your challenge is that on the $20 GPT plan you get a fairly good amount of Codex use and image gen. If I used Krea, it would be because I fine tuned/built a workflow on top of it to do something GPT image couldn't/wouldn't do. Maybe there's a market for the cheapest "good enough" bulk image gen model (spam? idk what the applications would be), but personally I'd say you should focus on workflow add-ons/one click fine tuning/etc that design companies can use to create assets with their custom "look" since the GPT-Image look will become blase from overuse.
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?
I think the early models distorted the perception of what artists actually need. It's not about workflows. Image creation is inherently iterative. None of the platforms seem to understand that, while AI artists have somehow fallen in love with the unnecessary complexity of node-based tools. Apparently, 400 tangled noodles look impressive when the boss peers over your shoulder.
The big studios and agencies are all building custom tooling as they always have but now a lot faster.
I've seen a graphical designer struggle to use an analogous product. They tried prompts like "remove background from this photo of dancers, don't change their face": the model couldn't do it: it always changed the faces at least a bit. In other words, the model was not removing the background, instead it was generating the whole picture anew, but without the background, which not only seemed inefficient, but was also way too lossy.
I wonder if this model could finish that task.
I wonder if your friend would have been more successful by removing "don't change their face" from the prompt
A lot of models required exactly that. But I imagine it varies from model to model.
Why do all the "frontier" models still occasionally output images with the horrible ai texture tells? I'm surprised whoever made the front page let those ones in lol. My hypothesis is that it is either an unpatching or feature loss shortcut/artifact. Until then I have no hopes of ai generated images being viable for anything actually beautiful or creative
Ex. https://optim-images.krea.ai/https---s-krea-ai-k3-presets-20...
A lot of the recent image models have that weird noise/dithering/subsampling issue like that. It's very noticeable.
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
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.
> 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!
(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.
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~
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!
> but at a fraction of the speed and a fraction of the price
"at a fraction of the speed" is generally understood to mean that it is much slower, but I don't think that is what you are intending to say (considering the title and website claim it is multiple times faster)?
oops. great catch. fixing rn!
I love krea 2, thank you for all your hard work!
megacool stuff
Comparison samples with GPT-Image-2.5 for UI design work:
Starship landing page
Krea 3: https://image.non.io/80b85aaa-286b-4ad7-b3ae-cfab4b491216.we...
GPT-Image-2.5: https://image.non.io/19c5f032-cbc7-4034-9682-3aba451bd669.we...
Fashion history website
Krea 3: https://image.non.io/a6e61219-963c-4921-b453-738960e03308.we...
GPT-Image-2.5: https://image.non.io/c067fd76-9c7e-46c2-a8fd-0b0931bf93f6.we...
Banking website
Krea 3: https://image.non.io/3a1e7b50-36e7-418c-ab47-368ce4f01948.we...
GPT-Image-2.5: https://image.non.io/b96a7e00-3289-4044-90de-e0f86a2ca55e.we...
Overall Krea 3 isn't near GPT-Image-2.5 in terms of visual creativity in this space, but it still should be commended for its ability to render small text details. That said despite it specifying you can add exact hex codes as color reference, I've found that they tend to leak into the design (check the valuation panel of the banking website - notice to the right of the "fixed indexed 7-yr" there's a hex code that made it into the image).
I'd love to get my hands on this to post-train and see if it can be pushed further in this space. I actually see a lot of potential for the model, despite the clear differences in the examples above.
That Banking website is a good example of GPT-Image-2.5 having absolutely no self control, to a point I am just shocked they havent been working to fix it. It's like the first thing the model does is "generate 100 slogans" and then try and figure out how to cram them all into the image.
You see it constantly, and its worse with GPT than Gemini models, but they both do it. They seem to think they need to tell not show. They cant not annotate their thoughts into the image. A lot of the time you see either random signs on posts added, tshirts with text describing the story, or hand written annotation over the image.
There's really no reason this should be a problem anymore on default settings, or that everybody should still be getting the exact same style of edgy poster out of GPT with icons for bullets, and lists that add nothing jammed into the image. They need to have some kind of seed library of templates they start from, and higher temperature on selecting different styles.
A reasonable conclusion, but you don't have all the data here in this particular instance. These are all part of a set of benchmark prompts I have that are designed to try and push the images to their extremes. The banking one is dense because the prompt itself specifies each of these things - the goal with this test is to see how well it can handle many different UI elements. Here's an excerpt from the json blob that is used as the prompt:
That's not to say that GPT-Image-2.5 doesn't have a certain smell to it though. It definitely has its isms.Not sure that UI design is necessarily the ideal or even recommended use case for image models.
I very firmly disagree. IMO image models are extremely under-used in this area.
Just for context, I ran design systems at Figma for 5 years / led their first prompt->design stream before quitting to make my own startup in this space.
Image models are trained on visual creativity, color usage, whitespace, etc - all the ingredients you need for good design principles. Where they struggle is often with information density and text rendering, which LLMs excel at. You can use the best parts of both of them though - use an image model for the initial design and an LLM to convert it to code.
One thing that's very telling is with openai's sites feature, in its chain of thought it actually renders out an image first as a design reference, then builds towards that.
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.
At least Krea doesn't have the GPT-image shimmer! All of the GPT-image images are immediately recognizable as AI.
Or the gpt-image "fried" texture. Interestingly the best approach I've found so far is to use gpt-image for the initial design, but nano banana for the asset export once I have the design. IE in the case of the "Annui" design, I'd take that design and its backing json structure, hand it off to Opus to build it in html, then use nano banana to build out the images of the statues/background images.
You just gotta use the best parts of all of them :)
I'm unfortunately getting the gpt-image fried thing: https://imgur.com/OYY1MSa
Which is interesting... after gpt-image-2 came out with this bug replacing the piss yellow filter, it seems like every other release except Google's have the same issue.
Off the top of my head, this exact artifact appeared in Qwen, Grok, and now Krea.
It makes me wonder if gpt-image-2 is being used to bootstrap post-training, or if their images are just so ubiquitous it's turned into a sort of "virus".
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
> Also, sometimes you want to do “ugly” things on purpose or add imperfections like film grain and such
This is pretty trivial with an extra sentence added to your prompt, for all the models. Getting them to make normal looking people is trickier.
As in that they don’t look like fashion models?
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
> 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
Not yet, but it's what I quit my role at Figma to make a bet on for my startup. In my experience diffusion is a far, far better creative designer than LLMs. LLMs are trained for consistent output, which makes sense for a billing function, but leads to some very obvious smell when doing frontend work (take claude design as an example - it's often very obvious what comes out of it). Image models have a bit more creativity baked in, and are trained on color usage, whitespace, and visual balance a lot more. I've found the image->html approach leads to much better results, basically using an image model as an architect.
Would love to be connected with the team for when you all are ready to finetune it. My email is j@diffui.ai
When VL models started becoming popular a while back, a relatively common workflow was to use generative models (like the original Nano Banana) to generate initial UI mockups and then throw the images at a VL model to convert them into CSS and HTML.
However feels like LLMs such as Opus 5.5 in particular are surprisingly adept at generating UI mockups right off the bat these days. And since they’re already code, they’re much easier to modify and tweak.
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.
I feel the same. For small motion graphics I am using claude in code and I think I didnt open AE for weeks. Bt when you need to put real image do scene (for example background or character) and it has to look same in 5-10 shots. Can krea keep it in same style/position through all edits? Or it can drift a bit?
The NCC-1701-D really is the default starship in our culture and the training corpus, isn't it.
Thank you. This shows that Krea 3 website isn't being honest, and this is what I hate about all those AI tools: the cherrypicking. Stop wasting our time!
And honestly, the whole thing looks like an ad-for-an-ad.
Are open weights being considered? The one thing I'm always looking forward to is all of the anime fine-tunes.
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!
Can you share any vague details on how much compute/train time was needed to create the model?
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!
2 images to try and then the only option after is a monthly subscription.
I don’t mind buying credits but I don’t want a subscription.
You can turn any subscription to a single purchase with a virtual credit card.
No.
In the USA, if you "subscribe" to something, that is a promise to pay on the regular, until you cancel or unsubscribe. It's not a "single purchase".
If you cancel your card, or use a virtual one that's immediately invalidated, the service will have every right to come after you, and bill you for subsequent installments. You will run up a bill. You can choose not to pay it, but you cannot say that you don't owe it, because you agreed to pay the bills in perpetuity until cancellation.
I have no idea why people claim this, and think that playing a shell game with their finances will stop it.
I’m sorry, but you are the one spreading misinformation.
But I’ll give you an easy way to prove that what you said is true instead of just yelling at each other.
Give me a single example in the US of an online service company providing service after a credit card declined on a no-commitment service with auto renewal that actually sued in court and recovered something. There is none.
Sneaky pricing. Offers you plans when you sign up, if you decline them, it offers you a discount.
Just offer the discounted price up front. If your business cannot survive without sneaky pricing, it's not a good business.
I detest this kind of pricing. It seems that for every subscription I need to threaten to leave to get a discount, and now need an AI to automate that process.
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.
> 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.
As a serial tech entrepreneur with a few decades of hands-on experience both in my own startups and also spending some years as a VP / product group owner inside an F500 acquirer, you may want to think more deeply on this. A hidden downside of modern A/B testing analytics is how easy they make it to min/max short-term gains while masking delayed and second-order effects.
In this instance, no doubt the analytics show this bumps ASP in the short-term but those analytics can't show you the downstream impacts. These tend to accrue over time creating an overall drag on the business that you have no way to identify. A good case in point is how Adobe successfully min/maxxed their way to a large portion of their customers either not trusting (or hating) them. The brand damage is now likely irreparable. Do what you must to make payroll but you're on a slippery slope. Be careful you don't let 'any dark pattern is okay, as long as it's legal and works' become status quo. The consequences can accrue faster than you can exit.
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.
Worry not, we have enough bandwidth to hate both the players and the game.
Ngl that made me laugh
The problems you desribe are real and you should talk more about them in general, but 'don't hate the player hate the game' always comes off as a flippant dismissal. You have a good product, I think you can do better than that.
That's a bullshit excuse for this particular issue. No one is forcing you to employ this specific dark pattern regardless of how you claim your competition behaves.
Great step in the right direction. IMO still to expensive for most “product” use cases.
Fast and affordable imaging editing unlocks tons of cool use cases. It’s refreshing to see latency be held just as important as price as-well!
I’ve been bootstrapping a personal styling app. Funnily enough I’m still seeing best results/$ from grok imagine V1. A now deprecated model.
The pearl earring and the billboard worker didn't feel like frontier quality. Otherwise very impressive
Gave it a photo of a woman wearing a T-shirt, asked it to seat her at a fancy bar. It invented a suit jacket, which makes sense, but the drawing on the shirt runs right over the jacket. Gemini did much better.
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.
I understand what you're saying. My first thought is "is that really applicable to a mistake like the text on the T-shirt extending over the suit jacket that partially covers it?" Maybe it is. I also don't know what you actually get when you prompt Krea 3 via their website as I did today. I don't know whether that folds in some guidance from an LLM first. Tough to say.
Why is there no comparison with FLUX 3 Image? I am not a big user of image models so maybe those are in different categories (or maybe it's not a relevant frontier model?). Just curious.
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.
Thank you for your answer
one thought : flux is open weight and available; krea is 'self-hosted capable with enterprise license agreement, contact us...' in the FAQ.
Maybe they don't want to make potential customers aware of competitive alternatives.
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!
Krea3 open-weights will be a legendary release. For me Krea2 was already one of those unforgettable drops. Same like Minimax H3 or Qwen3.8-27B.
Any relation to Krita?
"cre-" is a latin root meaning (more or less) "make" which is where we get "create" - both are likely drawing inspiration from that (as would be the case for Creo, a common CAD software).
The signage text on the fuzzy characters scene was scrambled. It's a no-go for me.
For *image* editing.
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”
Who told it to add that ugly shadow for the red car?
This looks really cool
Meh ... there are tons of great paid image models ... where the open weights at!?
they’ll come!
Good to know, hope it winds up on model garden soon
Maybe silly question but what’s model garden?
its the service in Google Cloud for accessing third party AI models hosted in Google’s data centers, alongside Google’s first party models
it also has enterprise level considerations that aren't present in direct API access of Google’s models, rendering almost all of the go to API documentation irrelevant, so it just comes from experience to know to use model garden’s implementation
Oh! We should put it there then!
I was aware of AWS Bedrock