Anything v3, BloodOrangeMix, and overfitting
#53
by
GiBi
- opened
I understand that when a model is"overfitted", it adheres too much to the data it was trained on, such that it cannot "generalize" or generate new data based on my prompts. A model I often hear people talk about being overfitted is Anything v3, which BloodOrangeMix is based on. I have a couple of questions...
- Did merging Anything v3 with other models solve the "overfitting" issue due to additional data being added to the model? In the model description for BloodOrangeMix, it is mentioned that it is "stable", what does this mean exactly?
- As an end user, should I really care about whether BloodOrangeMix (or any other OrangeMix model for this matter) is overfitted if all I want is a general purpose anime model that could do anything including adult content?
Hello @GiBi ,
While I can't speak to all of the details, I'll share what I understand.
You can find recipes used to create the models further down at https://huggingface.co/WarriorMama777/OrangeMixs#bloodorangemix-bom.
- They merged some NAI back in, and for the more NSFW and hardcore versions, added in gape60. Adding back in more NAI probably would help reduce the overfitting to the style you see very obviously in Anything v3. Not sure what is meant by stable, perhaps without distortions and with consistent results?
- No, all that matters is if you like the results, overfitting being bad entirely depends on what you are looking for result wise.
Stable means (was) stably popular.