⬅️V5 Luci and V4.9 Luci~➡️

Little Luci-chan comic down below!

TL;DR (SHORT VERSION) -

V5 update - New style, higher accuracy and no issues with microdetails (moles and more), new poses in the pipeline. What will get better? Find out below ⬇️

Set sizes and scope - Even close to 2 months later, NAI hasn't succeeded in finding the available compute to allow for more generations. Some adjustments had to be made and some features are dormant. Estimated set size going forward ~400-500 images.More info below ⬇️

Finetuning - Minor issues that'll be adjusted as we go. Won't take long to stabilize.

V5 and its capabilities? - Frontier level quality, high skill ceiling. ⬇️👀

Daily posting uninterrupted for 631 days! 🏆But will it continue? 🤔❓Will find out soon! 💪

Addressing the gooner nation! 🥸🤵

V5 update ✨

ℹ️ Fun little detail - V5 has accrued over 10,000 (surviving) generated test images with multitudes more in isolated A/B upscale tests, an estimated ~300 unique test environments with 35 days of logged active test days. And that's only based off existing data that wasn't deleted. 🤯

I understand that some may reminisce about the old look; just as you have grown accustomed to it, so have I. However, just like our transition from V3 to V4 (an era we might prefer to forget) and then into V4.5, it is now time to boldly step forward into the next age of advanced goonery.🫂

V5 Pros & Cons

PROS ✅:

Database cutoff - Early estimates put it to around ~June 2026 (That's a lot of new gacha girls 😈)

Visual aesthetic and consistency - The new style is very visually expressive and very colorful. Since the upstream latent image is cleaner that means that there's also going to be more stability per character as well.

Visual intelligence - Generation accuracy is extremely high, new poses will be introduced at a much faster pace.

Visual accuracy - Many of you may have noticed I've generally avoided characters with moles or other such details, well no more!

Male texture improvements - Males will look cooler and fit better into the...activites~

CONS❌:

Usage limits - Compute constraints make it harder to expand the sets vertically. Though I'm switching the design philosophy a bit as well.

Expect some growing pains - With a new model come the usual early hiccups, won't take long to fix everything on a prompt level.

Not much else - Model's great, they cooked. 😤

KNOWN ISSUES⏰:

Male skin texturing slightly underperforming - Will require some prompt balancing, won't nudge it too much tbh.

Textures can be a bit too glossy and shiny at times - stylistic leftover, again, not too sure if I even want it gone. It's tied to the lighting to an extent...

Male genitalia size too large - Prompt level adjustment, we like the meat 🍆 big here, but too big can compromise compositional freedom, so I'll tone it down with time.

Eye detailer slightly underperforming on mid-long range images - Will be fixed with tomorrows set.

Fingers can underperform - Will be alleviated with future sets on a prompt level

What's the model feel like to use and is it that good?🔬

The upstream latent source image is now the actual breadwinner of the sets, which means that not only will there be less upscaling related errors, but also characters will retain their unique identity far easier. As mentioned above, I've been shying away from characters will complicated microdetails like moles especially, the new upscaler architecture has been generally consistent, but the SDXL eye detailers were dropping the ball consistently. Though obviously complicated angles made thins unenjoyable as well. Long story short, that's basically been remedied almost completely. I've streamlined the pose creation process significantly as well and as soon as the first week of V5 is done and there's no substantial immediate issues, I'll move into significant horizontal expansion.

V5 has a substantially higher skill ceiling. It's not as easy to use out the gate and I had to employ quite a few tricks to stabilize outputs in a way I'd want them. Poses have been translated from V4.9 -> V5, but due to the new model training and post-training, they're not landing 100% of the time. Bugfixes are landing with each set.

What's the issue with NAI and the new content? 😟

NAI like any other AI company is fighting tooth and nail to secure cheap compute and inference. 💵💰💲

Competitors have more money and are eating everything up. 🐋To use a more consumer friendly example - An RTX 5090 now costs almost $10,000. It's very clear that AI is not slowing down and the big fish are getting bigger. NAI have done a good job as is and are alleviating the situation as best they can. The market share for hobby image generation is dwindling. Not doomposting or worrying too much though. NAI are still operating on a hefty profit margin, but the landscape around them is full of megalodons with billions of dollars to burn.

How that affects us? Due to the more stringent usage limits, I've flattened the top-end of sets to a more reasonable amount. They were on a slight downward trend beforehand already (intended as well).The current target is to bring them back to that ~400-500 range where sets still feel hefty and are still manageable.Now it's not like a 1000 image set was a "real" 1000 image set, the bulk of the images were simply alternate angles and viewports of an existing pose, but still, a cut is a cut (and for that I apologize🙇‍♂️). Poses will still be represented, but the variability per pose will now be seen across different characters, rather than within the set.

The design philosophy before was the more the better, in a way that's true, but at some point it comes at the cost of other things. In a good way, the time I've spent on the existing poses has trained me to basically write perfect prompts for what I want. That system has also been passed down to AI agents as well which allows for a much more streamlined process. The new design philosophy is to isolate an experience for every set that matches the character and their circumstances. The fabled themed sets now have a place in the system. Character archetypes will also play a role into how the poses are handled and "received"~. This system is currently dormant, but once the early V5 bugfixing is done, it'll quickly become noticeable. Themed sets will aim to focus on character and outfit specific traits. (You best bet your ass if a girl has a sleeveless armpit exposed dress she's getting the armpit treatment).

What about the visuals? 🎨

I think most people will enjoy the new style. It's more of a step back into the 2D domain with a focus on a stronger color palette and contrast, while also allowing for dynamic lighting that fills the image purposefully. The reality is the current anime stable diffusion technology is not capable of never-ending vertical progression in terms of visual quality, the weaker your latent, the more you have to upscale, the more you have to upscale, the further you get away from the latent. There's no way to cheat the stochastic approximation models, even with deterministic samplers. Noise is the law.So that means, move away from chasing the perfect textures, go back to 2D! RETURN TO TRADITION!

Posting schedule and my personal life 🗓️

Ties back to the usage limits mostly. But it's also due to my busier lifestyle as well. We're approaching the 2 year mark soon 🎇 and things have changed since we first started the journey.

I've been fortunate enough to find success here with every single one of you supporting me along, giving me ideas, giving feedback, and I want to extend a big ole' heartfelt Thank You! 🤗💝

I've also been fortunate to find a bit of success in my personal life as well, I've relocated with a new job offer and work is going great. The relocation was a rather substantial ordeal and took a lot of time which came at the cost of the models development and upkeep. In order to find a balance between the daily sets, my day-job, and developing new features, I could potentially miss a daily set every now and then.

As of right now it doesn't seem like it it'll happen, but just dropping it here that our award-winning streak might get interrupted unexpectedly. For the most part I can guarantee there won't be 2 days in a row without content. (If there is I got in a car accident or something).

Both working on and working with V5 has been tremendous fun, which is why I'd avoid calling it fatigue, since I genuinely enjoy doing what I do here. There is an element of accrued physical exhaustion, stress, and bad sleeping habits that can catch up to me regardless of how much I enjoy looking at big anime girl tits. I hope everyone can understand. 🫂

And a short little Luci-chan (in V5!) comic as usual~

Luci-chan unveiling V5 at the Oscars!!! And then she woke up...