V4.8 Update is Live! New poses, expanding PoVs, better textures, better backgrounds, and more!

Luci-chan doesn't like zoomer brainrot memes. 😔

Yes this is a comic of my daily married life as Luci-chan's househusband. High-Res Download link below~

Mostly the same technology, but expanded the scope and scalability.

TL;DR (SHORT VERSION) -

V4.8 update and the end of the focus on quality(maybe?)- 20 new poses (first batch), 2nd batch later this week, adjustments to existing poses, new multi-sampler approach.

New PoV's - Throwing in some female PoV's, experimenting with limited body poses (lower body only, head/upper body only)

Sets becoming massive - Sets are getting too big, the time for a switch to themed sets will begin. Ass, Feet, Armpits? Less messy, more messy? Heavy, low, or no NTR?

Dialogue problem cracked? - Local offline agent dedicated to customizing hentai plot? Is it possible?

Finetuning - As usual, each time a new model gets dropped there's small imperfections. Will seek to stabilize them asap.

Femboys delayed - ...again? Just a week or so please forgive me. 🙇‍♂️

New watermark - Sleeker and smaller.

Surely no one is gonna read all this yapping now, will they?

V4.8 update

What's new?

Multi-sampler approach.

Fancy nerd term.
Basically what happens is previously the image was generated by NAI as a baseline, thrown into an USDU 2x resolution upscaler, then passed through some face and eye ADetailers.

While that's cool and all, why not really tap into some power?
So that's exactly what I did. I upgraded my rig recently and I'm looking to really utilize every CUDA core available to me. After the NAI baseline image now we go through a 3-step multi-pass KSampler approach.
What does that mean exactly?
It means that the image is actually being holistically regenerated (albeit at a lower denoise level) at a encoded latent noise level as it passes through each KSampler that prioritizes different parts to get reconstructed.

What?
Long story short, there's more hardware power used on the entire image now, not just the eyes or face.

The 2x upscaling is now also done prematurely before it passes into the new system meaning that details are also handled on a 2x resolution as well. This roughly translates into clearer and finer small details (stray hairs, nail polish, jewelry, etc.).

2x this 2x that, why not 4K?
Long story short the actual upscaler models aren't designed to work at 4K. Also the upscaling time becomes prohibitively long. Nearly everyone is at either at 2K or within the error tolerances (~20% extra), but especially at 600+ image sets too many images will get sacrificed due to upscaling errors making it not worth the tradeoff.

The visual quality question.

I love working on making things look better and crisper.
But I think with this update I've finally hit diminishing returns. There's no other way at least to my knowledge or something that doesn't involve upgrading to an NVIDIA RTX PRO 6000 Blackwell GPU, to get the visual quality up higher.

At the end of the day this is all done off the back of NovelAI. And while I love the versatility they offer, at this point with their baseline (V4.5) version, this is as far as it can go. Unfortunately there's no news on V5 yet, hopefully it'll be sometime this year, but no guarantees.

While I won't stop fiddling with the knobs here and there, the main devtime will now switch from a neverending Sisyphean struggle for the most pristine pixels to expanding the content horizontally rather than vertically.

Which areas are going to be affected by the new V4.8 update?

1. Character textures.
   Expect to see more natural looking bodies. Less body shine (that slightly oily look sometimes), better female genetalia primarily, less shiny hair, better female pubic hair and general hair quality, nail polish, etc.

2. Backgrounds details.
   As the KSamplers are designed to work holistically with the entire image, there's going to be some effective background reconstruction. The full extent of this is something that will require substantially more observation to fully quantify. But expect to see uplift in bedsheet textures, background drawers, lamps and stationary lights, etc. All that kind of stuff.

3. Liquids.
   Slight improvement here again. Liquids look more 3D.

4. Clothing and other textures.
   Nothing too crazy here, just slightly more detailed fabrics.

5. Expressions effects.
   Expressions such as the disgusted expression will now have a more anime style effect around the eyes.

6. Softer makeup and lipstick.
   This is a bit of a balancing act. I'll keep an eye on it to gauge exactly how prominent or subdued it'll be under different circumstances.

New PoV's?

As many of you may have noticed near the end of the sets I recently introduced a lower body only pose. That's been expanded upon a little bit and once a bit more error proofing goes through I'll drop in a few more. The idea is to build a more rigid system allowing or truly "up close and personal" poses.

Female PoV -
Kind of a similar concept. This pose will exist in limited capacity during the "main" image rotation, with themed sets enjoying more instances of this PoV. While I'm caling it Female PoV, it'd be more accurate to describe it as a "Taker PoV" as this will also work with Femboy sets.

Close-ups -
This one is a bit harder as I need to expand the clothing appending system itself. Right now the limiting factor are the body group classifications I have. tl;dr there's no distinction between the above cleavage area and the above navel area, all of that would be considered "upper body", or the midriff area and below the pubic area, all of that would be considered "lower body". I recently finalized a lot of front-facing and backside data, but the next step would be this.

Sets becoming massive

Both v4.8 demo sets are hiting 750+ images.
Obviously I designed them to be as error-free as possible with signature clothing and a simple background, but that doesn't change the fact that these numbers are getting a bit crazy. While I've been slashing the total generative amount, the error rate is so low that I'm enjoying the leisure of handpicking only the best responding images.

The solution? Themed sets really.
This has been in the pipeline for a while, but with the recent hardware upgrade I got caught up in trying to utilize it. Themed sets are nothing new, they're one of the many mythic things I talk about from time to time and then conveniently forget whenever I had to drop a new update (don't punish me...). But finally as mentioned before, the focus will be on horizontal growth in variety, rather than vertically in visual clarity.

What themes are we looking at?
Initially they'll be simple.

The first layer of themed focus. (1️⃣ First Update).
More ass focused, more feet focused, more armpit focused, etc.

Second layer will focus on the roughness of the intercourse and other poses.(2️⃣ Second update)
A lot of the groundwork here has already been laid for this, but basically just more runny makeup, more sloppier saliva, more messy hair, stronger ecstasy expressions, etc.
This will maybe be less binary and more of a randomized feature with a coverage slider. Meaning some sets will have a 20% "messiness" rate to them, some an 80%. That kind of thing.

A third layer that I'm planning on is the level of NTR content. ( 3️⃣Third update)
Light NTR sets will focus on more "loving" NTR themes (holding the viewers hand during some poses, dialogue is more focused on reasurring and being thankful, apprehensive/shy to the events occurring, etc.)
Heavy NTR sets will focus on more "mocking/disgusted" NTR themes (verbal ridicule, slapping the viewers hand away, spitting on the viewer, dialogue is more focused on mocking and humiliating, confident/disgusted to events occurring, etc.)
No/Limited NTR sets will focus on a situation where the viewer is not directly involved in the occurring NTR. (think for example a recorded video or something similar, basically kind of what we have now).

What's the ratio at which these individual themed sets will occur?
Initially, since the first and second upate are fairly palatable to pretty much everyone. Those will rotate nearly all the time.
The third update will be the bulk of the work naturally and will most likely have fairly restricted usage initially as to maintain the novelty.

Sets will have a main pose progression which will account for about 60-70% (haha the meme) of the volume of a set. Themed content will take up the rest at a reduced image count.
Example -
Set A (No NTR) - 600 images
Set B (Heavy NTR) - 400 baseline image progression, 100 themed images

These numbers are pure estimations, but that's the idea.

Dialogue problem cracked?

What was the problem with the dialogue feature?
It basically felt too repetitive and manually expanding and curating it was cumbersome and difficult to those as uncreative as me.

The solution? A local agent.
I'm sure everyone's been bombarded with news about all the advancements in local-ran AI. Everything from the Mac Mini hype, Deepseek, OpenClaw and now mini Qwen models. Models have become so lightweight and quantized that it's very possible to have it run in the background virtually all the time.

I've done some preliminary exploration and I have my very own Luci-chan OpenClaw agent running in the background. She helps me with a lot of things (the paid requests, the comments you guys send and she's helping me automate a lot of the social media stuff).
The idea is I'm going to offload the dialogue system to her and have her work on creating character-specific dialogue scenarios.
Think Hinata Hyuuga from Naruto, Luci-chan is gonna gonna take into account the shows synopsis and provide multiple scripts for potential NTR scenarios (maybe one where she's still not married to Naruto, maybe one after their marriage, etc.)

Ambitious. But will it work, and if, when?
Preliminary tests work so far and the idea behind the OpenClaw agent is that it has the capability to actually write and save these scripts autonomously, as well as message me and keep me up to date with those scripts via Telegram or Discord.
Though this will come AFTER the third update is finalized with themed sets. While it's not super far away, it's not soon either. The important thing is that the technical foundation is there.

Finetuning

Known issues -

1. Eyes are slightly smaller than before (about 10-20%). Maybe a bit too small at some angles.
   Will be fixed with the next set.

2. Fat rolls sometimes or extra belly folds.
   Addressed on a prompt level, will monitor how substantial this is. I don't necessarily dislike it, but it can be a bit questionable for some characters.

3. Extra navel.
   Can happen in the case of clothing texture splitting or overlapping the navel area. Will manually clean them up until I figure out some alternative solution. Apologies if I miss any.

4. Upside down shots cause errors in the eye detailer.
   WIll be alleviated with the next set.

Femboys delayed. AGAIN???

Bro trust me bro... I love femboys with all my heart just as much as the next straight guy.

I just simply didn't have enough time to bundle them up with this update. The upscaler took 12 FULL revisions, not even including the per cycle refinements. Femboys are completely functional on a prompt level, I simply haven't had the time to generate a full-fledged set and edit it.

When...
Literally like next weekend (unless something goes wrong).

Smaller watermark.

Is it smaller?
I don't actually know. Seems like it? More neutral color and a more stylish font at least.

That's all for now.

A pretty big update overall. Definitely a bit more finetuning necessary than before. The eyes specifically are going to require a lot of monitoring to make sure they're perfect. I've fixed my sleep schedule recently and my day job has been reasonable lately. Rolling up my sleeves and getting back to work. 🫡