I would assume it will be as soon as they are able. Last I heard, they were having some issues getting costs down and proper optimizations in place for the text-to-image model. They also said that it could be ready as soon as April, assuming everything went smoothly with the text-to-image upgrade.

Essentially, sit tight. It will be done when it's done.

First, you will need a powerful PC and GPU. AI is very resource hungry and takes a lot of processing power to be very good. This is especially true if you want it to "remember" things.

It's using Chroma which is a custom version of Flux.1-Schnell. I believe it's one of the recent snapshots of training data which somewhere around ~40 of 50.

I could see you being upset if perchance was some paid service with deadlines and shit. But, it's not. It's free. What do you even have to be upset about? Besides you're own entertainment, what do you even have invested on perchance?

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.

It is possible (and quite standard) to add additional training data to a model after deployment. It takes a long time, and a considerable amount of GPU power. It is also the only way we are likely to get good celebrity support again, because the vast majority of models are now shipped without training on celebrity data sets anymore.

Golly gee, dev. I know you're providing a completely free service and ads barely cover provided services, but go ahead and add these extra features that will raise costs further.

It's free! What the hell? 🙄

Well, I did see an em-dash in there...

Anyway, I do understand the sentiment. It's not very fun being left in the dark, especially when we're not even given a chance to participate.

I'm willing to bet the training data for those tags simply hasn't completed and/or hasn't started yet.

Quote from the notes you can find by clicking the little link that appears when you are generating an image:

This new image model is still training. It's currently worse than the previous one at some things. It'll steadily improve in quality and gain new knowledge over the next few months.

MONTHS

Training takes a long time and a lot of the GPU power on the hardware this model runs on (part of the speed issue). It is getting better as more training data completes, but please be patient.

That's... strange. Normally it's the other way around, unless you are doing something unusual, since let has block scope, and var has function scope. Those error normally come from using var as the variable then get's "hoisted" and so becomes visible before the variable is actually declared.

Anyway, good that you were able to figure it out.

Because AI text gen is not just spitting out text from some lookup table: it's looking through billions of chunks of text, looking for patterns, then combining them together to form a response. There is a an absolutely massive amount of number crunching that goes into the process. In addition, your want for it to "remember" things is an extremely tall order because, AI doesn't remember anything.

Every query you send (this is true for all AI models) includes:

• role instructions
• previous messages
• "memories" (blocks of text with summarized context)
• your query

This is always sent as one big blob of text that the AI has to read, parse, process, group, compare, quantize, then finally generate a response. The more data you send, the more it has to process, the more resources are required.

Your idea to write to the reminder of a character seems viable. Perhaps just a collection of keywords in the reminder? Then, each character bio could contain a description + context on what those keywords mean?

Another option could be something similar to the Strict Game Master that uses code to track things like inventory and skills.

Yet another could be leveraging /mem. That is in chronological order (I think) so maybe you can add entries there?

Weren't you the one that made the post that this is in response to?

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.

Quote from the notes you can find by clicking the little link that appears when you are generating an image:

This new image model is still training. It's currently worse than the previous one at some things. It'll steadily improve in quality and gain new knowledge over the next few months.

MONTHS

Training takes a long time and a lot of the GPU power on the hardware this model runs on (part of the speed issue). It is getting better as more training data completes, but please be patient.

In other words: you are not imagining things, it will get better. Hopefully everyone can just calm down and be patient.

You're not a customer. You're a user of a free service provided as part of a larger service.

What do you mean? It's useful information.

The site? Image gen is just a plugin on the site. It's a site for creating generators, which are small programs and pages for creating various randomized content. Image gen is a plugin that can be imported into a generator.

I don't understand how changing one plugin has ruined the whole site.

From what I can gather, the dev doesn't like interacting very much. This is something that they do in their spare time as a fulltime dev. They created the perchance platform first and foremost as a way for people to learn code and web development. The AI was always a side feature.

There has also been a lot of grumbling in the perchance community (both here, in the comments of various generators, on reddit, and on the Discord server) about the quality of the previous image generator (plus plenty of feedback via the feedback tool, I'm sure), so it isn't really a surprise that he/she finally made the switch to a new model after stating that they planned to in more than one post here in this forum.

In addition, I didn't see anyone (before the change in model) praising the old one... only complaints about the quality, resolution, coherence, etc. Now that a new model is in place, suddenly many people have lots good things to say about the old one.

Bottom line, this is a completely free, uncensored, private, and unlimited service that the dev was kind enough to gift to us. I'm certainly not going to complain when I don't even know what the final outcome is going to be yet.

You know, a lot of the trouble could have been avoided if people would simply read the notes about the change and have a little patience. It says right in there that the model is still training, so it literally doesn't have all the data needed to make images similar to the old one, yet. It's starting at base settings and parameters, so expect a lot of strange variance and "sameness".

For those wondering why we can't have the old one as an option still, the answer is simple: memory (RAM). These image models don't spin up instantly, so they need to remain loaded in memory in order to field requests. To have the old one running at the same time would mean effectively doubling the resources required when the dev is already running on a shoestring budget supported entirely by ad revenue. It's either one running or the other. In order to train the new model, it must be running, which means the old one can't run at the same time.