Comunity.

What if instead of working separately on creating loras models, we joined together in small groups?

What if instead of focusing on creating loras in an integrated, sequential, and individualized way, we divided the process into small and simple parts for people with less technical knowledge? Supporting each other and speeding up the model creation process.

This post complements and gives meaning to the protocol I presented to the world last month: How to train MORE LORA models in LESS time and in a simpler way? As some comments mentioned, and even a reviewer accused me of being "all talk," this is not a major technical change. I confirm this in this post, it is not meant to be. I have no idea how to make loras faster in that sense, but I can contribute speed and improvements in everything else. Let's be honest, there are very important bottlenecks in lora creation at an organizational level, many of which could be improved/solved with measures like this.

<aside> 🆕 Summary or "Abstract" in general lines for lazy people, since some complain about the amount of text I put in my publications xD: (all this eliminates the examples, fancy marketing, and big words and explanations that I leave below)

• A training laboratory is a server where different people coordinate or share resources to create more or better models.
• Training laboratories solve many current problems in model creation (See Why is a training laboratory useful?)
• We have to create more than one, with different moral rules to make it work without problems since reaching an agreement among everyone is very idealistic. (See How are we going to do it?))
• Anyone who wants can share theirs in the comments of this post, we will add them to the post as time passes and they are trustworthy.
• I leave some links and example resources of what a training laboratory could be, its presentation, rules, ways of doing things. All the links I share are from the training laboratory that I will launch soon

If you like the initiative and even want to participate, I encourage you to keep reading.

</aside>

Why is a training lab useful?

Two heads are better than one.

So far, Civitai has followed a clear structure of volunteers who create datasets, train and test loras on their own., However, this process has many problems. For example, the model creation ends up following a single line without feedback during the process or collaborative organization.Some creators have realized this and have added forms to request models, to have initial feedback on what models they could create. However, this is not a scalable way of doing things, as it still depends on a single creator. If for some reason that person has to leave their tasks, the model would be left orphaned. Another problem with this method is that if the trainer has a question, they cannot directly contact the dataset requester. These are just some of the multiple bottleneck problems that would be solved with a place that connects people more quickly and organized:

I'll give you another example. One person can create loras that other people may altruistically want, and that's great. But that person will always have a problem, which is that they will have to create datasets for characters they might not know. Therefore, they could make mistakes (e.g., accidentally adding another similar character to the dataset). This could be solved in multiple ways, such as allowing dataset creation processes to be done by the requesters themselves, as it is relatively simple. In this way, people with less advanced equipment but who know the characters could contribute to the model creation. Doesn't that seem like reason enough? Let's move on to another example: model testing is a tedious task that must be done to see if a lora has turned out well or badly. This, like datasets, could be done by third parties willing to run tests. And well, if I keep giving more examples, it will get dark. Do you get it? Of course, I would like to point out that all these examples are not new, I'm not claiming to be the inventor of this idea. Something similar already existed in servers like Civitai or Waifu Diffusion, which have sections with more trainer/user feedback. So, what would be the difference between future training labs and Discord servers with sections to request models?

Facil, a diferencia de los servidores generales que ya existen y tienen secciones así en discord, un laboratorio de entrenamiento, gira completamente en torno a la creación de modelos, con opciones como roles personalizados para organizar funciones y fases (Creación de dataser, prueba de modelos, entrenador de modelos…), secciones separadas, tutoriales especificos anclados, personas realmente comprometidas a hacer mejores modelos y más cosas.

Easy. Unlike the general servers that already exist and have sections like this on Discord, a training lab revolves completely around the creation of models. It offers options like custom roles to organize functions and phases (dataset creation, model testing, model training, etc.), separate sections, pinned specific tutorials, and people who are truly committed to creating better models and more things.

Simile to understand the power of this idea:

Inspiration from the AI audio community: