New (small) V4.6c Update! Visual update targeting eyes and facial improvements! (Patreon)

New V4.6c Update!

Not really worthy of a big monthly update post, but I might as well put it here instead of bloating the set posts.

The update mainly focuses on facial clarity improvements. ✨Roughly an increase of around ~20%-40% to eye clarity and fidelity.

I introduced a few more enhancement nodes to the upscaling system to allow for much more precise targetting of the eye area, as well as upgraded the existing node that handles face clarity.Improvements will also be noticeable in eyeshadow and general makeup, as well as visual effects such as runny makeup, tears and symbol-shaped pupils. Visual effect improvements are not limited to just the eye area, as the mouth area effects will also benefit (saliva, teeth, other "liquids" in and around the mouth, etc.)

More indepth breakdown further below! ⬇️

The update will be most noticeable with characters with unorthodox pupils and more complicated eye textures. The pilot set Emilia from Re:Zero served as the launch pad for this update and has both good examples of V4.6c performance as well as some images that still require finetuning. Emilia has these peculiar slit-like brighter (or at the very least not dark) pupils as well as a spherical effect to her irises. Obviously this is a bit too much for the AI to recreate at all distances, but more often than not, it managed to at least incorporate some, if not all of the elements in some capacity in the close-up poses. Will be finetuned in time, but the idea is obviously since there's no tags for this, data from the character specific trained dataset is used to allow for this to happen.

Here's a visual representation of what's occurred.

On the left is the old V4.6b version, on the right the V4.6c. While the differences aren't substantial I'll outline the clear differences below.

For example let's break it down -

The lipstick in this time around wasn't that strikingly different, but from now on it will be properly applied to both the lower and upper lip. Color saturation should be better as well. I'll monitor as it doesn't become excessive, or I'll maybe start picking the makeup colors rather than letting them freeform. I do like variety in the makeup personally, but again, I'll try to make it not excessive. This should also affect teeth quality and clarity, as well as saliva and maybe make mouth expressions a bit more accurate and crisp.

Thick eyelashes are properly thick now.I like the extra thick eyelashes A LOT, so this will probably not be reduced in any capacity xd.

More accurate shadows that fall over the face.The new nodes should focus on accuracy like that especially. While hard to notice, the base image doesn't properly cast a shadow from her blunt bangs and has these extra shadows for strands of hair that don't exist.

The pupils should be more accurately applied within the iris. The idea is to minimize the empty space that occurs in the original image where there's a gap between the bright pupil and the outer line of the pupil.

While not as noticeable, the lower lash line and eye contours will see slight visual effects akin to Korean style aegyo sal makeup. While not strictly applied to all images, it's an effect that in results from random generation that do have it, the new enhancement nodes should allow for it to be properly visualized.

Not really worth pointing out but small differences in the shading, blush, and stray hair texturing may be seen. While not as important, I'll be monitoring this over time as well. But yes, all of those should see improvements over time as well.

That's just a small and simple breakdown on a random good responder image. As usual, not all images will benefit fully, some may error out a bit and/or overproduce effects.

I still physically inspect and pass every single image before putting it in the oven, so I'll try to always cross-select for good responders. So far I've not noticed any drastic changes in the error rate (which has been successfully brought down to around 1% since the launch of V4.6!). But of course I appreciate everyone's patience with the ongoing enhancements, and even this will slowly with time be almost completely bugfixed.

Another example, the V4.6b preparing for Christmas Luci-chan ( sleepy ).

And the V4.6c preparing for Christmas Luci-chan ( sleepy, but her makeup didn't get ruined by the pillow ).

Here the differences aren't as noticeable in the iris area, but the makeup definitely looks much more pronounced. As well it's noticeable that the excess moles under the eyes have been cleaned up, the lipstick has become slightly more pronounced and the stray hairs are textured slightly better.

That's all for now.

While this was a smaller update, it's actually a pretty substantial one since it allows for instant integration with the future V5 update from NovelAI. While there's no news on it yet, I'm expecting maybe a sneak peek or something around the holiday season (cope).