It's interesting to hear people side with the artists when in previous discussions on this forum I've gotten significant approval/agreement arguing that copyright is far too long. As I've argued in the past, I think copyright should last maybe five years: in this modern era, monetizing your work doesn't (usually) have to take more than a short time. I'd happily concede to some sort of renewal process to extend that p…
I thought the US wants to re-industrialize now? Then 5 years is laughably short to protect your invest.
Yes. They should look for a job that actually covers their lifestyle instead of crying about AI taking their jobs. They can always put in more hours and become better. I cant imagine they have a lot of paying customers anyway.
I keep waiting for physical objects to become important again. AI isn't coming for the ceramic folks.
> AI isn't coming for the ceramic folks.
*looks at 3D printer on desk that can apparently handle ceramic filaments, thinks about all the mass-produced ceramics sold in supermarkets*
4o still exhibits the "pink elephant effect", it's just... subtler, and tends to reveal itself on a complex or confusing prompt. Negations are also still not handled properly, they tend to slightly confuse the model and decrease the accuracy of the answer or the generated picture. The same is true for any other LLM. Moreover, the author is asking the model to rationalize the decision he already made ("tell me why the…
But it's literally a different architecture (auto-regressive, presumably sequence based vs diffusion). In my experiments it is significantly, overwhelmingly better at consistency, coherence and prompt adherence. Things I needed control nets before it just... does it. and even zooming into fine details, they make sense. Here is an example with a bunch of negations: https://i.imgur.com/P8G5ICs.png
It’s interesting that out of all the aquatic animals it could have used, it chose one that perhaps looks the most like an elephant.
This is a before/after moment for image generation. A simple example is the background images on a ton of (mediocre) music youtube channels. They almost all use AI generated images that are full of nonsense the closer you look. Jazz channels will feature coffee shops with garbled text on the menu and furniture blending together. I bet all of that disappears over the next few months. On another note, and perhaps other…
> They almost all use AI generated images that are full of nonsense the closer you look. Jazz channels will feature coffee shops with garbled text on the menu and furniture blending together.
Noticed tht.
Maybe it's my algorithm but YouTube is seemingly filled with these videos now.
But it's literally a different architecture (auto-regressive, presumably sequence based vs diffusion). In my experiments it is significantly, overwhelmingly better at consistency, coherence and prompt adherence. Things I needed control nets before it just... does it. and even zooming into fine details, they make sense. Here is an example with a bunch of negations: https://i.imgur.com/P8G5ICs.png
It’s interesting that out of all the aquatic animals it could have used, it chose one that perhaps looks the most like an elephant.
Wha- wha- what? I tried to generate an image in ChatGPT after the announcement a while back and the image wasn't bad, but the text on it (numbers) was nonsense. (Analog gauge with nonsense numbers instead of e.g. 10, 20, 30, 40, etc.) Gave it another chance now, explicitly calling out the numbers. Well, they are improved but not sure how useful this result is (the spacing between numbers is a little off and there's s…
These images are still created with the old model. The share link states "Made with the old version of image generation. New images coming soon." below the first image.
4o still exhibits the "pink elephant effect", it's just... subtler, and tends to reveal itself on a complex or confusing prompt. Negations are also still not handled properly, they tend to slightly confuse the model and decrease the accuracy of the answer or the generated picture. The same is true for any other LLM. Moreover, the author is asking the model to rationalize the decision he already made ("tell me why the…
But it's literally a different architecture (auto-regressive, presumably sequence based vs diffusion). In my experiments it is significantly, overwhelmingly better at consistency, coherence and prompt adherence. Things I needed control nets before it just... does it. and even zooming into fine details, they make sense. Here is an example with a bunch of negations: https://i.imgur.com/P8G5ICs.png
Of course, it's a tiny specialized model vs a big generalist model. They're absolutely incomparable in size/quality though, especially of the text part. How much of this happens because of the poor encoder and worse training in other models, and how much due to the architectural differences? I'm not saying it's not better than the existing image gen models somehow, but it's pretty hard to separate the two since both are present. All current SotA LLMs including 4o itself have negation inaccuracy in text (you need a really complex prompt, or a long one with thousands of tokens, not a toy one), and I don't see why this one should behave differently in similar conditions. Especially considering that it also suffers from pretty much the same artifacts as other image models, just much less (fingers, extra limbs, perspective/lighting issues, overfitting, struggles with out-of-distribution generation etc.)
I'm all for paying artists for their work. Unfortunately, same as tattoo artists, some just heavily overcharge for mediocre results (been tattooing myself AND I know a few things about art). Like, sorry, but if you want to earn money doing art, please be good at it...
> some just heavily overcharge for mediocre results if people are paying, then they aren't "overcharging"
In the tattoo business people have no other places to go. Charging 1k€+ for half a sleeve is extremely overcharged. If people are paying they often simply don't have enough alternatives.
Huh, the coffee table reminds me of all those cheap e-retailers who very clearly (and badly) photoshop their clothes on to same 2 or 3 stock model images. If you thought shopping online sucked before, it's just going to get even worse now.
Looking at the example where the coffee table is swapped, I notice every time the image is reprocessed it mutates, based on the previous iteration, and objects become more bizarre each time, like chinese whispers. * The weird-ass basket decoration on the table originally has some big chain links (maybe anchor chain, to keep the theme with the beach painting). By the third version, they're leathery and are merging wit…
It's kind of clear that for every request, it generates a new image entirely. Some people are speculating a diffusion decoder but i think it's more likely an implementation of VAR - https://arxiv.org/abs/2404.02905.
So rather than predicting each patch at the target resolution right away, it starts with the image (as patches) at a very small resolution and increasingly scales up. I guess that could make it hard for the model to learn to just copy and paste image tokens for editing like it might for text.
