I thought OpenAI had been sleeping a bit and given up on the image generation race. However, with the recent release of the oddly named 4o Image Generation model (Why not continue with the DALL-E naming scheme? I personally love the pun) it looks like they cooked up one hell of a model.
If you are into visual GenAI you have probably already seen many examples of quite incredible outputs from the new model. However, we decided that we wanted to make a large scale evaluation, based on 200k human responses across 13k image pairings.
Unfortunately that also meant that we had to generate a large amount of new images, and since OpenAI have not yet opened up API access, we had to do it manually through the UI :(.
The benchmark tests the model in coherence, prompt-alignment, and overall aesthetic preference. Especially for the first two, OpenAI's new model is very far ahead of the competition.
Check out the detailed results and the collected data which is openly available on huggingface!
Cool! Did not know about this. I was not able to see anything other than the pipe - which I assume is the SDR versions, maybe some clearer instructions/limitations (and description of what you should expect to see) would be beneficial.
Pretty cool! And also really difficult as it turns out. Love that you can try it without signing up!
Runs quite smooth and I like the simplicity of the UI. Maybe a button to return to the starting position is missing?
Call me old fashioned, but I think a speech should be something personal that reflects you and the person/people on the receiving end of the speech. If you do not know what to say, then maybe there is no need for a speech.
That being said, it could be helpful to some people that have something particular things on their mind, but unsure how to structure it into a speech.
Did not really manage to get very far yet, but maybe I will revisit later. I was very confused by frequently going to that landing page, which is still WIP, but I guess it will make more sense to me if I spend some more time.
In general I was missing a bit of intro, but I assume that is part of the experience
Is there a reason you did not submit the thread with the link? I was a little bit confused that the title did not take me to the link and I think most HN users would expect that as well - just a pointer :)
Sounds really cool, but i agree with the previous comment that some example would be a very cool, as cloning and running the repo is a bit cumbersome to quickly check it out.
What are your hopes with this? Is it just for yourself to use, or do you hope to get a large gallery of submissions?
I was hoping I could submit to the gallery without logging in, but in retrospect I do see why that probably should not be allowed
I agree that in my experience, AI image generation does generally still struggle a lot with generating nice simplistic graphics relevant for these types of uses. However, I think the idea of having a large enough catalogue of graphics, pictograms, and backgrounds that an "AI" could choose from based on the instructions would give a much more customized experience.
I like the idea of a design tool that helps making nice looking infographics. Initially I assumed that the graphics would also be generated, or at least be looked up in a relatively extensive database. However, it appears that the template dictates the layout and the only thing that is really generated is the text, which imo does not warrant a specific tool.
In the end, I think this tool is a little misleading currently and could be a lot more than what it currently is