Luci-chan wishes you all a Happy 4th of July! 🎉

Little Luci-chan comic down below!

TL;DR (SHORT VERSION) -

V4.9 update - 40 new poses (some still a bit buggy), adjustments to existing poses. New upscaling architecture (goodbye Illustrious SDXL👋).
What will get better? Find out below ⬇️

Set sizes - Managed to barely squeeze in the 40 new poses into this by cutting a lot of the old content that wasn't very good and refining timings on underachievers and overpeformers. New content from here on has to sacrifice existing poses.

Dynamic framing - Patched around half of the total poses to be able to handle dynamic framing (from below, from side, from above, etc.)

Dialogue in trouble? - Due to the significant architectural shift, dialogue can't survive past the current upscaling pass. As usual, I'll research ways to do it.

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

Updated artist mix + prompt adjustments & VFX - Reframed a lot of poses to include the male in the image more!

The future of the models and V5? - This year. ⬇️👀

Now onto what's going on~

V4.9 update ✨

ℹ️ Fun little detail - V4.9 took an isolated 104 test environments to get finalized.

What has improved?

V4.8c

↕️

V4.9

PROS ✅ in just this image:

• Better jewelry (pendants).

• More accurate clothing (central part of the leotard, sleeves, thigh boots).

• Better hair texture.

• Better obfuscated details (earrings).

• Less chaotic backgrounds.

• Better fingers (error fixed + textured better as well).

• More moody lighting.

• Less random sweating.

• Slightly more mature/deeper look to the eyes.

• More accurate shadows.

• Better clothing texture (bunny ears).

• Object permanence (her right hand behind her backside).

• Better ambient occlusion (near creased regions like the armpits and between the legs).

• Softer, safer body textures (less duplicate navels or aggressive skin creasing around the abdomen).

CONS❌ in general:

• Not visible here, but male body in general is softer (not necessarily bad).

• Slight visual deviation on more aggressive expressions (Not bad either, just a more expressive model).

• The more aggressive the model is the more violent its mistakes will be. I manually inspect every single set before I set it off to upload, but sometimes I'm sleepy eepy and/or watching a stream on my 3rd screen (need a 4th tbh) and can miss a bad image. I apologize for that 🙇‍♂️

What's so good about this new architecture?🔬

Well... long story short, the previous model architecture was just outdated. Illustrious-based UNet's and CLIP text encoders are just a thing of the past. SDXL's 4-channel VAE fails on fine details like fingers and other overlapping meshes. The cutoff was like 2024(?) for characters which means any upscaling was it just guessing randomly and patching fixes to make characters not bug out was tedious. That means that micro details on clothing, hair ornaments, and signature details will see massive improvement. No more random micro details like moles and inaccurate navels and skin creases!

What could get improved with time?📈

This is an actual new technology that's in active development. Every update previously was me battling with something that hasn't been updated since 2024 (late 2023 tbh). That means the actual Stable Diffusion lab will put out updates and patches which will improve the visual quality without me having to jump through hoops to figure some frankenstein fix.

Set Sizes.⭕

Currently I'm at the literal possible maximum. I cut down 15 or so bad poses and swapped in 40 new ones. Sizes as usual will get squished down to around 750ish images. New poses still need a bit of time to mature and get fully bugfixed. Give it a week or so and you'll be seeing them regularly. After this any new addition will come at the cost of legacy poses or underachievers. I've made some new tools to keep track of it and bugfixing has been streamlined.

Dynamic Framing.🎦

Nothing too crazy about this. Simply ran all the existing poses through a visual inspection on how far they can bend until they break. Expect the current legacy poses to differ slightly with new from above, from below, from side and such perspectives!

Updated artist mix and other adjustments & VFX.✨

Basically reframed a few of the poses to allow for more of the male to become visible in a lot of the poses. I'm going to increase that a little bit with time as well. The idea is to allow for better utilization of screenspace. Older versions kind of gave up too much space to the girl. Slight downside is the males are back to being Mr. Olympia winners and the genitalia size is slightly larger. Will finetune!

VFX are going to get expanded on pretty soon. SFX like panting, plapping, Ah-ing, etc. Other effects such as skin ripple (booty shake), and such are now possible (they weren't before due to how sensitive the skin was), more messy hair, messy liquids, etc.

V5 when and what's the future lookin' like? 👀

This year.
V5 is in active training and will be released by NAI sometime after post training, eval, and finetuning is done.

As for the new architecture.
The lab is active, the community is picking it up. The outlook is good. This is a great load off my shoulders because Illustrious based SDXL was a dead-end, the team sold out and they've been embroiled in drama for years now. Arguably no one's been adapting it past 0.1/1.0, so the scene was dead outside of finetuning and LoRA training. Now, don't get me wrong those aren't underwhelming endeavors, but when you have an actual research lab working actively and spending significant amounts of money on the entire training cycle, that just makes me hopeful for future updates.
Without a doubt the strongest update will be V5, and with NAI actively training that means that we're past the halfway mark.

Anything else?

Just a bit of venting -

• Adding new poses.
  The workflow for adding new poses has improved drastically, I can basically add a new pose in like less than 10 minutes now. The issue is actually finding them. For the 40 new poses I actually went through 500 pages of Danbooru art (with filters).

• The technical difficulty and devtime.
  The sophisticated upscaling is the active roadblock that's preventing me from expanding more narrative-based contnet. I had a working pilot earlier last month for a dynamic storytelling system that'd spin up small preludes and epilogues for the characters based on the location the set is taking place in and a few other things, but the clash with the upscaler made it so that I had to pick a lane to devote devtime to. Luckily the new architecture is capable of handling text, it's just a matter of time before I update the system to accomodate it.
  
  It's not an understatement to say this update took somewhere in the ballpark of ~300 active hours of devtime to finalize. On one of my days off I ended up working 17 hours debugging a floating point imprecision error that was causing NaN's due to --highvram loadsettings. Right now there's an active bug with NVIDIA's v13 CUDA where something poisons cached VRAM in between architecture handoffs and it's what slowed down the update so much. Regardless, I adapted and built around that issue, but that's all just devtime that doesn't produce new content. There were over 104 individualized testing environments made for the update (as in a customized ComfyUI workflow to test different parameters all at once).

  I was really close to missing a day of posting especially in the last couple of days, but I managed to keep the streak! It's been an uninterrupted 596 day streak (I think).

I'm thankful for everyone's patience with this update.🙇‍♂️

It unfortunately took a little bit longer than anticipated and I apologize for failing to deliver on things you may have been hopeful for.
I'll continue to try to meet everyones expectations, especially over the next ~2.5 months! I have a nice little lull in my day job over that period of time, so I'll try to catch up as best I can on complex and exciting new content!

That's it from me for now, I gotta go enjoy some barbecue. 👋

Little 4th of July Luci-chan comic in the new V4.9 look!

Luci-chan got caught vibecoding again 💢