Manything, a PoC

This model is a simple proof of concept with one objective in mind: to validate if it's possible to fix the problem of the "same face" that plages most SD models.

The name is a sort of a pun that comes from the base model used for this test, Anithing and a reference to the idea of trying to make the model able to produce many faces.

The problem

The "same face" problem is well known and can be seen in many models of all kinds of styles. After models have been merged with each other for so long, it has become very difficult to find models that don't produce the face of the same girl, only with different wigs and costumes, unless you use black magic to force the model to produce anything different.

To complicate further, the face has spread to many concepts that if used automatically make the models produce this face. For example:

1girl, girl, woman, female, human, person, realistic, hyperrealistic, photo, photorealistic, masterpiece, best quality, ...

And additionally, it has also been tied to essential parts of the images, such as shadows, composition and others, making it not trivial to fix.

In some cases the influence of the sameface can be so extreme, that it becomes noticeable even when generating male characters.

When using a model without this issue, normally the faces produced will be inferred based on the rest of the prompt, usually adapting to the context provided, without having to force the model in that direction.

The test

A relatively simple way the problem could be fixed with is to identify which layers of the model are responsible for the "sameface" and swap those with a model without the problem using weighted block merge (WBM), but this changes the behavior of the model into something different.

Instead, a relatively simple iterative process to remove sameface and maintain other characteristics of the model was used.

The result

While it was not possible to maintain the rest of the model identical to the original while removing the sameface issue, an initial result of the test does produce more varied faces more easily and the influence of the sameface girl is greatly reduced. More importantly, the faces have characteristics and proportions that are closer to an average human instead of the easily recognizable "sameface" AI face. The rest of the prompt will have more influence on the faces generated.

All of the images in the showcase use the same seed and negative prompt, to compare both model under the same conditions.

There may or may not be more versions if there are additional tests, at the moment I only wanted to see if it was possible. The model should work using the same settings as the original model.