CreamyClayK wrote:so how does the AI handle pictures from other angles? most of the subjects in the pictures are facing the camera straight on or with a slight angle, but nothing from behind. does the AI not handle it well or are there just not enough source materials to go off of?
AI handles them quite well. Here's a render of QSA's "Sinkhole":
I don't do much in terms of rear-view / butt sinks, but AI is perfectly capable of working with them. I've seen other AI artists exclusively produce butt angles.
sixgunzloaded wrote:So, these models - how are they made? How is such realism achieved? I noticed these last 3 characters all look very similar. Do they use a 3d figure with a morphable mesh, or are they, like, generated from an accumulated library of real faces and physiques? I've read that these AI companies pull from sources all over the internet, referencing both photographic and cgi imagery. (Many times without permission from the original sources. But that's a different discussion.)
Just fascinating because they are so life-like compared to any polygonal figures I've come across, except maybe for Project MetaHuman for Unreal Engine 5.
All gorgeous images too, by the way! Thanks!
It's important to clarify that a "model" in AI doesn't refer to the character (as in, a Poser model). A "model" is the algorithm used to generate the product. In simple terms, it's the data set used (among other factors). The NovelAI model is mostly an anime-based data set and its prompts and tags are based of anime databases. Something like Stable Diffusion has a broader data set and is more general-purpose.
The "fine-tuned" models are subsets. Because SD is so broad, the results are almost unworkable since what you want is very specific (e.g. a digital illustration, a photograph, etc.). People therefore create these fine-tuned models to draw upon a narrower range of samples to get a more predictable outcome. For example, one fine-tuned model might be much better suited to getting the style and shape right for a cartoon look. Another might be better suited for a realistic look, Another might be more suited for Asians, or males, or furries, and so on. Compare the following, where I ran the same prompt and image reference through three different fine-tuned models:
These are all 2D images. It doesn't use a 3D model. What it does is that it "rolls" each part of the image to approximate what it calculates is the desired output. Imagine you threw buckets of paint at a wall until it turns into a painting. That's basically what a diffusion model does. Its training set allows to recognise the best patterns to throw paint in to get the best result.
The resemblance of the three red-heads is partially coincidental, but also partially because I used a similar set of source images and didn't do much to transform the face shapes. AI tends to put its renders into an "uncanny valley" in creating similar human faces.
The high levels of realism are derived from choosing the right fine-tuned models that are better suited to photorealistic outputs, drawing upon a narrower set of realistic samples. I've been using Stable Diffusion - Realistic Vision V2, but the recent release of LifeLike was the reason why I tried this set. Someone else should be able to replicate this output by using my exact prompts and SD model.
