GrumblePuss

What do mean? Which part is the bad omen?

Perhaps it's two celebrities in the same prompt? The assumption may be that if they are both there, then you may be depicting those famous people as doing something together, which could be a big red flag.

The training is happening on the server side, by the dev. Many model creators refuse to use celebrities and famous people as part of the training data (with a common exception for public figures) for their models. However, there is no rule about creating databases of these images and then using them to train models that you host. That's what the training is for, to fill in the gaps that are in the model.

Refer to these notes if you want some more info.

What you are referring to is known as "training". Training is the process of showing the AI model thousands upon thousands of images. These images come from curated datasets. Each image is accompanied by some text that describes in clear terms exactly what is in each image. The AI model analyzes each image, compares the image to the text, then stores the data for later use. The more images and text, the more the model will understand. The more the model understands, the better it will be able to use the data it has.

Currently, the model is in training. As training continues you will see the image generator steadily improve until it understands what the old model did.

Perhaps my analogy wasn't very good?

Flux is a dramatically better image gen model. Most consider it better than SDXL, yet still lightweight enough (with a bunch of optimization) to run on the limited resources the dev can afford. When the training is complete, this model will be far more capable then the old one ever was, but the training needs to finish for it to get better.

I'm sorry you are having trouble with this, but a large number of people have complained about the quality of the old model, yet there is no way to fix those issues without upgrading the model. Now that the old model is being replaced, all many people can do is complain, rather than be thankful that the dev is working hard to give us a better model.

To use a loose analogy, that's like getting a new (much better) car and calling it worse than the previous one because the fuel tank hasn't been filled with gas yet.

Training is the bread and butter that makes an image generator work. It is the reference data that it pulls from to create images. Without it, there is no "fuel" for the engine to run, so the results will be sub-optimal.

And, again, training is an extremely GPU intensive process. The fact the dev is able to run the training data and still support image gen at a usable level is pure wizardry.

Of course it isn't a coincidence. That is the entire point of the training! Training will literally improve the output quality as it is fed more data on different styles, celebrities, artists, etc. this model does not include everything, and that was for a reason. The dev is attempting to run this model on very limited resources, so they loaded the base model (not to mention, many of the popular/new models are deliberately excluding data on celebs for fear of lawsuits). So, it's going to need training because it needs to be very carefully trained in order to keep memory and GPU usage down.

So, no, you are not going to make the hardware work faster.

How do you expect both to be hosted? That would be double the resource usage, double the cost. Not to mention, the actual training (which is still running and will continue for the next month or so) will dramatically improve results and consistency. It simply doesn't have all it's data yet. Until then, it's going to have some issues, but it will get much better.

Ah, thanks. People have been so angry about the new model lately. I'm just trying to get this information out there.

What do you mean? It's useful information.