I evaluated Tess.design about a year ago for an app I was building. At first I was excited because I wanted a service that compensated artists. However the number of artists was very limited and the blog post said “more will be added soon” but it had already been a year and it seemed like none had been added, not a good sign.
Then I tested out the image generation itself and I was unable to come up with prompts that achieved the kind of images I wanted. My only prior experience at the time was OpenAI API. With OpenAI I usually got what I wanted on the first or second try, but with Tess, I couldn’t get a usable result even after 20 tries.
So in addition to the limited number of artists, I think the quality of outputs vs. competing models was a huge factor. I needed to generate thousands of images, so I couldn’t afford to do dozens of attempts for each one.
Hopefully one day there will be a service that can match the quality of OpenAI Image API and Flux but with compensation for artists.
Yeah this just shows that ergonomics matters. I use Nano Banana and Grok Imagine to generate silly images for my friends and siblings (instead of reaction gifs I do reaction slop). The workflow is quite easy. Just plop in a prompt and usually the first image is good enough to share. Not that my standards are high anyway.
Would I pay extra to ensure that the artists that these models were trained on were compensated fairly? Absolutely! Would I pay extra for that but with degraded ergonomics? Given that this is just a silly hobby, probably not, if I'm being honest.
I think if that problem can be solved, and it's marketed to the correct group, a player in this space could certainly do well.
Most people can't even imagine the complexity it would require to actually build a system that correctly tracks down the sources for image generation. Not to mention that each image is generated from literally every single training image in a very small percentage.
It's not hard when someone inputs "create in style of studio ghibli" to say that studio Ghibli should get a cut. It's very different when you don't specify the source for the origin style.
And if you tried to identify the source material owner, the percentage of the output image that their work contributed to would be extremely - if not infinitely - small. You'd get minuscule payouts.
Most people can't even imagine the complexity required to build an LLM. But here we are.
We throw plenty of smart people at plenty of hard ideas. If a company really wanted to, they would find a way to make this feasible.
Funny thing is that building an LLM isn't as complex as you might think.
But the problem of attribution is easily understandable to any human with a modicum of intelligence.
Imagine that you have a trillion input images, with every single one having a source associated. When training they go through lots of processes and every single image contributes a varying degree to a subset of 8billion parameters. That alone would produce a dataset that is 1T * 8B to just say how much a particular image contributed to the output...
To mimic intelligence the output is also randomized - the association is not static and every single pixel in the output has it's own lineage.
So as you can probably imagine that to calculate the actual source weights on the output you'd require to do at least 8e+21 calculations per output pixel... and require double precision floating point while you do it.
We know how to do it. It's just ridiculously expensive.
(The above example is for demonstrative purposes only)
Insults aside, you chose a very expensive way to solve the attribution problem. But my rebuttal is simple: we are literally commenting in a thread about an AI image generator that paid people. It didn't work, but if a company I've never heard of can try an experiment like this, I'm sure our billion dollar AI overlords could if they wanted to.