I'm getting mixed results, and for a given topic it seems to invariably give a better result first time you ask, then not so good if you ask again.
It could be random and my imagination, but seems that way.
Show HN: New AI edits images based on text instructions
51–60 of 241 posts
Re: Show HN: New AI edits images based on text instructions
#52The headline and the heavy promotional verbiage on the site seems to be claiming this is some new functionality we didn’t have before. Image2image with text instructions isn’t new as the headline implies.
InvokeAI (and a few other projects as well) already does all this stuff much better unless I’m missing something. There are plenty of stable diffusion wrappers. Why not help improve them instead of copying them?
I’m not against having enthusiasm for one’s project, but tell us why this is different and please don’t pretend the other projects don’t have this stuff.
Re: Show HN: New AI edits images based on text instructions
#53It's a little premature, fine, but I want to start liquidating my rhetorical swaps here: I've been saying since last summer (sometimes on HN, sometimes elsewhere) that "prompt engineering" is BS and that in a world where AI gets better and better, expecting to develop lasting competency in an area of AI-adjacent performance (a.k.a. telling an AI what to do in exactly the right way to get the right result) is akin to…
Haven't we been here before? - see self driving cars.
I've certainly seen this argument before..
Yes, it's true that not all technology evolves as fast as predicted (by some, at some point), but first of all I still believe we will see self driving cars in the future and secondly, it's one anti-example in a forest of examples of tech that evolves beyond anyone's expectations. I don't find it very convincing.
Re: Show HN: New AI edits images based on text instructions
#54The headline and the heavy promotional verbiage on the site seems to be claiming this is some new functionality we didn’t have before. Image2image with text instructions isn’t new as the headline implies. InvokeAI (and a few other projects as well) already does all this stuff much better unless I’m missing something. There are plenty of stable diffusion wrappers. Why not help improve them instead of copying them? I’m…
I'm not aware of any pre-existing open-source model that selectively edits images (leaving some parts untouched) based on instructions. This new method is much better than the image2image that shipped with the original stable diffusion. I'm looking at the InvokeAI docs right now and don't see anything like this feature. We previously had smart-masks, but InstructPix2Pix mostly does away with the need for those as well.
If I am mistaken please provide links to these prior features.
Re: Show HN: New AI edits images based on text instructions
#55Can it make it pop? Because that was the #1 request I remember dealing with.
I don’t know why people use this “AI” thing, I have been using make my logo bigger cream (tm) for ages with success.
https://www.youtube.com/watch?v=GOwi3x92teo
;)
Re: Show HN: New AI edits images based on text instructions
#56So, Deckard can ask it to enhance, finally :)
Re: Show HN: New AI edits images based on text instructions
#57It's very interesting, thanks!
I've noticed (on the Spock example) that "make him smile" didn't produce a very... "comely" result (he basically becomes a vampire).
I was thinking of deploying something like that in one of our app features, but I'm scared of making our Users look like vampires :-)
Is it your experience that the model struggles more with faces than with other changes?
Re: Show HN: New AI edits images based on text instructions
#58Is there a link to how this works - in terms of nn architecture to combine the embedding of the existing image with the edit instruction?
Re: Show HN: New AI edits images based on text instructions
#59Where can I find more data about the work you did to create this?
Re: Show HN: New AI edits images based on text instructions
#60It's a little premature, fine, but I want to start liquidating my rhetorical swaps here: I've been saying since last summer (sometimes on HN, sometimes elsewhere) that "prompt engineering" is BS and that in a world where AI gets better and better, expecting to develop lasting competency in an area of AI-adjacent performance (a.k.a. telling an AI what to do in exactly the right way to get the right result) is akin to…
IMHO it stems from lack of imagination. Impressive as the results may sometimes be, the user interfaces for AI are still extremely crude.
Soon we will see AI being used to define semantic operations on images that are hard to define exactly (imagine a knob to make an image more or less "cyberpunk", for example).
I also expect AI-powered inpainting to become a ubiquitous piece of functionality in drawing and editing tools (there are already Photoshop plugins).
Furthermore, my hunch is that many of the use cases around image creation will gradually move towards direct manipulation. Somewhat like painting, but without a physical model. AI components will be probably applied to interpreting the user's touch input in a similar way to how they are currently deployed to understand text input.