Show HN: Web app to generate AI pictures with logos "hidden" in them

[igorkotua]

Hey there

Some time ago I built a site that generates AI pictures in a style of GitHub logo (https://octoart.vercel.app/). It's completely free, and even official GitHub accounts in Twitter and Instagram (yes turns out they have Instagram) posted about it. It was fun.

So I thought it would be cool to build an app that works with any logo, so people can generate nice branded pictures with their logo.

I went on to build https://logopictureai.com/. It's build with Next.js, Replicate API and Supabase.

It's a very early version, I built it in a weekend.

It works like this: your upload a logo, type a prompt (or select a predefined one), select number of variations to generate and click a button. Images will be delivered to your email in 2-3 minutes.

As I said, it's pretty early and I am not sure if it's really useful. It works decently with most logos, but sometimes can generate something weird.

Anyways, would love to hear your thought. Thanks!

[toddmorey]

Not meant as a criticism, just an observation: The novelty of these shapes hidden in images is already gone for me. It's amazing to think how fast the hype cycles around a clever concept will accelerate with generative AI being so available and easy.

Meanwhile if people still had to do this stuff by hand in Photoshop, you could probably milk the gimmick for at least a couple of years Thomas Kinkade style.

I'm reminded a bit of million dollar homepage and all the fast-follow clones that each earned about 86¢.

[jprete]

There’s also a lack of actual communication. Yes, an AI image generator can stick a pattern into an otherwise totally different kind of image. But there’s no communicative meaning there other than what I just mentioned. There’s no cleverness, nothing interesting to discover in the details, no space where the imagination can fill things in because it’s all plainly arbitrary, no deeper meaning other than a shape embedded in a pattern fill, however complicated the fill might be.

I went looking at the details of the original spiral medieval town and all I found was a woman’s skirt standing on its own with no body. And I know perfectly well the AI isn’t trying to say anything clever with that.

[toddmorey]

That’s true. Even when AI learns to not put torso-less legs, the wonder will be gone.

Prior to AI, you’d stare at an image like this and I believe most of the magic was marveling at the cleverness & craft of the person who painstakingly created the illusion. That you could think, “wow I’d never have the patience, talent, or dedication to make something like that.”

[solardev]

IMO there's a place in the world for easy, low-cost niche services that save you time.

I use things like remove.bg (a background remover) or VectorMagic (for raster vectorization) frequently, even though I have decades of experience in Adobe. Sometimes it's just easier/good enough to go through a dedicated tool that repackages some common workflow into a simpler wizard rather than have to DIY everything.

I tried to set up my own stable diffusion workflow the other day, using the ready-built GUI app Draw Things. Several hours of work and many gigabytes later, I managed to create several iterations of low-quality images that reminded me of early AI a few years ago. None of it approached the ease of use or output quality of, say, Midjourney, much less the state of the art these days (whatever it may be).

Maybe I could've tinkered with it for a few more days/weeks, but why? It's just an occasional toy hobby use for me, not something I want to get a degree in just to understand.

My main gripe with these services is their pricing model. I don't want to subscribe to something I might only use a few times a year. Wish I could just plop down like $10-$20 for a bucket of tokens that I can use-as-I-go. But that's not very sustainable for their business, I guess.

[willsmith72]

I think that's just about the context of where you see the content.

It's not that interesting as a hacker news post, or somewhere else with a tech audience.

The actual value for these things is in the mainstream, making content creation and graphic design easier.

In those fields, it's not about the novelty of, "wow that's impressive, how did they do that?". It's just about creating engaging content which is aesthetic and pleasing to people.

[digging]

I expect it to be less engaging/pleasing and more instant article-filler material. To actually make something artistically interesting still requires effort, just a different kind of effort. The vast majority of these sorts of images will probably be uninteresting gags and background art.

In a way, the same is true of photography and digital art already, so this isn't a total condemnation (that is, most pieces have almost no value). I just don't think this particular kind of image generation is going to be a vehicle for anything noteworthy.

[digging]

> The novelty of these shapes hidden in images is already gone for me.

Yeah I think it took approximately 1 AI image of cats spelling naughty words for me to crest that hill and lose interest in the entire technology. It's a technique that was only ever interesting because it took effort and cleverness, it turns out.

[ibdf]

This looks interesting, but honestly, for a tool that can create infinite amount of images it's odd that you only have 1 low res logo example on the right side side (which seems to come from your other project). It's not enough to tell what the capabilities of the tool are, and it's not enough to get me to pay for it. I would suggest either a free-trial (even if it's super short) or provide more examples with different type of logos.

FYI: all links on the bottom of your page lead to a 404. (Contact us Terms & Conditions Refund Policy Blog)

[igorkotua]

Yep, website is not polished yet. Definitely should add more pictures, agree

I can also add like ~5 free images to new accounts when they sign up. Adding more might get pricey for me.

Thanks for your thoughts

[solardev]

Side note: Just nitpicking, but on the topic of website polish, you might want to hire a copywriter (or an AI, I guess) to rephrase the English text on the page. It sounds very unnatural right now.

Example: "You can no longer puzzle over where to get beautiful pictures" makes it sound like "You can't solve this picture puzzle anymore", which isn't what you meant. Maybe something like "Stop worrying and let LogoPicture make you the perfect brand image" or "Make a beautiful photo with your logo", etc.

But that's just one of the many awkward phrases on the page.

[isoprophlex]

Also, several grating spelling mistakes. Even running the copy thru chatgpt would drastically improve matters.

[igorkotua]

Thanks for pointing out, yeah it was very fast paced

[xnx]

Here are some instructions for anyone who wants to do this on their own hardware: https://takin.ai/learn/generate-images-with-hidden-text-usin...

Definitely a nice convenience feature to have someone pre-select the settings, but very doable for anyone with tech experience to set up by themselves.

[broken_clock]

You can get something similar for free here:

https://huggingface.co/spaces/AP123/IllusionDiffusion

[serial_dev]

But can it write CSS so that the site looks acceptable on mobile???

Snarkiness aside, the website looks awful on mobile, half the screen is not used, overflow issues, etc.

[jsf01]

I think most of this would be resolved by just reducing the left and right padding on small screens. If the author is monitoring this thread, please consider making that update!

[igorkotua]

Just did, thank you! Sorry about that

[minimaxir]

The original submission is likely using a ControlNet implementation. You don't need the special control_v1p_sd15_qrcode_monster ControlNet that the popular optical illusions technique uses: the common canny and depth ControlNets works fine.

When ControlNet first became mainstream I generated a fun series of tech company logo charcuterie boards using the Canny ControlNet and SD 1.5: https://www.reddit.com/r/StableDiffusion/comments/11bp30o/te...

Of course now with SDXL you can create very good ControlNet-constrained images.

EDIT: In another comment the OP confirms it uses an optical illusion ControlNet implementation.

[teaearlgraycold]

You really need some kind of free-tier, but I understand you're putting yourself at risk with that. Ideally you could run the model in the browser with WebGPU for trial users so that they could bear the cost.

[all2]

I'd be curious about how big the model is. The loading time could be quite long. I suppose with caching, it won't be a big deal after the first run, though.

[andybak]

It's gonna be SD 1.5 and ControlNet, surely?

[wscourge]

This website looks real bad on my phone. Google Pixel 4a (5G).

[wepple]

Same on iPhone; clicked away immediately

[igorkotua]

Yeah sorry about that, it's not mobile optimized yet

[joenot443]

In 2023, when 60% of the world’s traffic comes from mobile, it’s not really an option I’m afraid. Not supporting mobile is like not supporting SSL at this point.

[igorkotua]

Yeah I agree, it's already fixed

[uticus]

I could see how this would be great for special event materials etc... if the pictures are what I want? There's no way to scroll through and choose, or even filter or change to get closer to something I want. So, no. Not interested in paying for a grab-box/gamble.

If there were a way to fine-tune, then pick 10 or something, it would be a lot more useful. Could even watermark or something to keep freeloaders away, hopefully you get the idea.

[emilfihlman]

Feedback on the site:

The English on the it is very bad and unnatural, and unfortunately it's right at the start

"You can no longer puzzle over where to get beautiful pictures for your brand. You can create optical illusion art with your logo in a few minutes!"

You really need to fix this!

Also consider adding a way to just wait and download the image from the website itself!

[igorkotua]

Thanks for your comment! Grammar will be fixed

Downloading pictures from the website will be added soon, just wanted to launch the thing as fast as possible

[Minor49er]

Looking at the "Answers to common questions about LogoPicture AI", the first question is "Do you offer refunds?" This makes me think that this is the question that most people are asking. Hopefully that isn't the case because this looks pretty neat

[aczerepinski]

Price feels fair for saving me time, but I have no use for a fairly low-res image. Do brands really want logo images that are less than 4k resolution and aren’t vectorized?

[qingcharles]

You have to run it through another AI app to upscale it. That's how most of these AI images are being done these days.

Vector is another matter. I don't think you'd want to try to vectorize the output from this, but again, there are a bunch of AI tools for doing it.

[aczerepinski]

If I have the skills and willingness to use another AI, then I might as well do the whole thing by myself.

Agreed on this not being the type of image you’d vectorize. In this specific case I’d want high res images suitable for print, or for a logo on a 4k video, etc.

[andybak]

How would these even vectorize?

[igorkotua]

There is a pretty good website and API for it: https://vectorizer.ai/

[andybak]

I didn't mean that. I meant these aren't the kind of images where you're going to get usable vectors.

[qingcharles]

Yeah, agreed, that's why I said you probably wouldn't want to do it in my comment, but you know someone will ask for it... :)

[igorkotua]

Yeah totally agree, it's just for the first version. Upscaling will be built into the app

[willsmith72]

I'd try it if there was a free tier.

[igorkotua]

Thinking about free tier

[quickthrower2]

The harder problem to solve is designing the logo in the first place

[offminded]

@levelsio is this you and what year is this? 2021?

You might want to get a rate limiter - expose yourself a tiny bit and offer a free tier before asking for money right away, it is a bit turn off for visitors.

But nevertheless bold move posting your replicate API based paywall website at HN first.

[nevodavid10]

What model in replicate did you use?

[igorkotua]

I use a model which is trained to generate QR codes, it works well with logo too

[omneity]

You can use vanilla sdxl with controlnet.

[porsager]

So the exact same as https://ailogoart.com ?

[constantly]

It doesn't matter. They're all just fungible wrappers on OpenAI or Midjourney or some open source model or whatever the flavor of the week is. I don't blame people for doing this -- those "generate your profile photo in 20 different scenes with AI" that were big six to nine months ago (?) I heard generated literal millions from people putting a frontend in front of whatever AI source before those style of images fell out of fashion.

The critical error of OP's attempt at this was not spending the barest amount of thought or time on designing the website or writing reasonable copy. That, and the number of businesses who will use this are much smaller than the population of regular people who want to do something like AI-ify their profile photos.

[CharlotteSophie]

Really cool!

[dvrp]

so, like krea ai?

[RosanaAnaDana]

The number of people hijacking foss tools and wrapping them up behind paywalls is hilarious.

[drusepth]

Really just shows how little effort is put into making those foss tools easy to use for the layperson (NB: not implying that they should, just that they don't). Turns out people will pay to be able to use a slightly simplified version of something cool, that they would otherwise need to "figure out" a bunch of stuff to get it working.

[tayo42]

Going from cool to usable for general public is a lot of work. And most of it isn't fun or rewarding. I get why open source doesn't do that, and why you would want money to go from solving a problem for your self to solving a problem for others in an easy way.

[solardev]

> Going from cool to usable for general public is a lot of work. And most of it isn't fun or rewarding.

As a frontend dev, this is my favorite part of the work! I wish there was an easy way to find open-source projects wanting UI help in particular. Is there?

[all2]

Oooh! Ooh! There's a guy who just posted a tool that has a limited license UI. He said he was going to work on rolling a new one. You should reach out maybe?

This guy! https://news.ycombinator.com/item?id=38044782

[kyleyeats]

You have to with this tech. If your thing takes off and you don't, you're stuck with a huge bill. Then you get leapfrogged anyway.

[burkaman]

I think it's the same as restaurants "hijacking" cheap ingredients and serving them up behind a paywall. Just because you think you could make a tool yourself doesn't mean it's never worth paying for, even professional chefs will pay for something pre-made and easy when they're hungry.

[garrrikkotua]

What do you mean hijacking?