Xenos-Camicature - Subtle Caricature and Art Style Realism Enhancer For Photo Real Images

Adds a slight caricature effect to your output for both a pretty cool style and for the reasons outlined below.

As a neat side effect - this model can take something heavily geared toward an Anime style face and push it back toward a more Western Art and/or Realistic looking shape. This is Galena REDUX with Camicature at 0, 0.7, and 1.0 weights.

Background & Reasoning

With AI Art, creating a consistent character can be tricky, at best. The easiest way to do that is to find a person or character known by your model (be it a checkpoint, LoRa, or whatever) and use that as your character's baseline.

Having grown up over the history of home computing, there was a time in the 90's when there was a little program called Photoshop that had evolved to the point where talented artists could superimpose celebrity faces on other images to put them in... well... erm... compromising positions or make it look like they were doing or promoting something that they never did. It was misleading at best and illegal in many situations too. And if nothing else, it was just plain creepy. I vowed early in life never to be a part of that.

With the advent of SD and AI Art, this issue is heating up again. When it comes to various art style models and such - it's not really a big deal. The image you are looking at is obviously an artistic representation of the person or character in question - not something being created to create the illusion that it actually is an image of said subject.

Alas, the photorealistic models are getting better - and they are a lot of fun to play with, too. And there are some things out there now that sort of push the boundaries of creepy infringement.

So... if I don't want to make realistic art that has consistent characters but doesn't come off as trying to actually BE a photo of that character? What do I do?

Camicature

Caricature art is when someone distorts and exaggerates certain features of a subject so that they are still recognizable - but it's obviously not a true and realistic representation of the person.

Camicature art is that same concept, but rather than it looking like someone drew it with a pencil, brush, or other art media tool - it still retains the look of being taken by a camera - i.e. There's still a sense of realism there. Well... at least mostly.

One of the challenges I had when trying to figure out how to do this was that if you use the term "caricature" in most models, it will tend to make it look drawn - you lose the photography/realism of it. That's fine when you're making something that represents a hand drawn art style (like my XenoEngine Model) - but it doesn't work so well when you're trying to represent realism and photographic output.

While working on that model, I spotted a few things happening and thought that once I was happy with that - I'd come back to those things I saw and try to refine them. This LoRa is the result of that.

What To Expect

This does, in fact, change the art style a bit. It won't look exactly like a simply (and slightly) exaggerated face. I had to start with the concept of a caricature (which makes it look drawn) and then push it backwards toward reality with a mix of other concepts. There may be a better and more effective way to do this - but at the skill level I have right now - it was the best approach I had.

Tips and Shortfalls:

• This worked well with the half dozen or so realistic (or realism capable) Checkpoints I tried it on. It also worked well on some of the art style models - giving it just a little extra (there is a Galena REDUX example or two in the promo images).

• This is (probably) not going to work well with models that mix in LoRa Styles by default (such as my XenoEngine model - really, it's BAD. That is, in part, the things Camicature is doing are already being done by the XenoEngine model - so it's basically just doubling up everything.)

• Having "Restore Faces" set to off is sometimes a better result. There are a lot of factors involved here - such as camera distance and the quality and style of the source images your subject was trained on. You'll need to experiment a bit.

• Playing with the weight of this can help. 1.0 looked great in almost all my trial runs. I got it up to around 1.5 before it started to burn in too much and look contrasty. And I could still see it's effect but at a lesser level down to around 0.5 weight.

• This is not designed to make heavy caricature distortions - just enough to make it obvious that this is a "representation of" the subject and not an actual photo of them.

• At certain angles, the face can sometimes go a bit weird (and the conditions vary a lot depending upon the subject - and probably the source images used to train the model on the concept of the person or character). Using the go to "distortion" type negative prompts can help in these situations. (e.g. distorted face, bad face, etc.)

Making Consistent Characters Bonus Tip

This trick is something many of you already know, but I see questions about it all the time on Reddit and other areas. So, I figured I'd share it here.