TeeKay's Titty Time Checkpoint Merge!
0.5beta GGUF Q4:
My first quantized version, so let me know if I screwed something up. I re-gen'd most of the 0.5beta sample images; there are slight differences, with some bigger than others, but for the most part not a huge difference.
0.5beta fp8:
Started from scratch with new updated LoRA's and then went through a few iterations of blending with 0.4beta version. This one shouldn't have as many occasional blurry gens. Cheers!
0.4beta fp8:
A bit of pussy added in order to counter the barbie doll smooth groin from before. If you're only aiming for topless, I'd still go with 0.3beta; this one adds pussy, but can lose a bit of definition sometimes. You may also still need a vag lora or post gen detailer; there was only so much I could add before I started to affect the tiddies and facial aesthetic.
Big thanks to Skullkid for his Pussy Diffusion LoRA, as this is what I used to add that bit of spice.
0.3beta fp8:
fp8 version added! I ran the same images loading all parameters using the fp8 model; the differences ranged from slight to near non-existent. Again, first time in this arena so let me know if something's not working. Thanks!
Original 0.3beta Description:
I've done a lot of lora's and versions of lora's, but this is my first go at a base model checkpoint. It's a merge with the full Flux1.D model, blending in a selection of my lora's to focus mainly on nude portraits. Nipples should come through well and the tits themselves should be a more natural shape with a bit of an upward turn; they're also going to be pretty big.
Like I said, this is my first try at a checkpoint, so all feedback is much appreciated. Does it do something well? How much does it do really shittily? (cuz I expect there to be more of the latter than the former 😅 )
Sampler: As far as settings go, I used Euler pretty much every time in my normal workflow, so it didn't even occur to me to test any others out. I'll update this area if I do some testing in the future.
Scheduler: I like using a custom scheduler, but I've found it messes with my metadata saves, so I now default to Normal Scheduler with 20-30 steps.
Description
fp8 version
FAQ
Comments (9)
So it's been a while. Can the latest fine tuned Flux models do full on NSFW or not? Because, it doesn't seem so. Or am I wrong?
if you mean can they do sexual acts out of the box, similar to old school porn merge models from sd1.5 days and pretty much all pony models, then no, I don't believe i've seen anything like that yet. Lora's are required and even then it can be a bear of a time to get them to work.
as for this model, it currently only does topless, my next step is to add in some vag, but we'll see how it goes, but I don't ever foresee it doing any hardcore stuff without outside help.
@TeeKay Over time perhaps? Despite all the attention toward the power of the Flux models, I seem to be using Pony models most of all. Surprising. But would you say Flux is still the best and most advanced and that over time it will overtake the more flexible models like Pony and Illustrious?
@olternaut I don't think it'll get to the same point as pony in terms of prompting positions, but I've been surprised before, so you never know. Flux sacrifices a lot of flexibility for realism and accuracy of facial structure/body parts/etc. Celeb lora's on Flux, for example, are a miracle compared to pony and predecessors. So it comes down to what your intention is, then you can pick the model that best suits the situation.
It has T5 sized roadblock. So unless someone figures out how to remove it, nope. Think you will see sooner AuraFlow based Pony v7.
@Mescalamba What does the roadblock entail? Just so I can fully understand what's going on.
@olternaut Quite simple, more hardcore concepts than nudes (it will refuse even more hardcore nudes, try forcing it naked chick with legs spread apart) it goes into model and around layer 5 or 6 (at least in case T5 XL), your NSFW hardcore concept is castrated/censored and continues on as much softer version, till is spit out neutered at last layer of T5 encoder, where its presumably eaten by FLUX.
I mean, even models that can in theory do for example one chick sucking nipple of other (meaning that model has learned it) cant actually pull it, cause T5 will prevent it.
Only somewhat reliable bypass is forcing CLIP over T5, one guy had some semi-success, but it seems its really hard to make such model work.
Think it would be easier to use some uncensored Gemma2 2B, modify it for Lumina 2.0 and train that as next PONY model. Or as PONY guy actually does, he uses completely unlimited version of T5 as instructor for AuraFlow 0.3 (which is going to be next PONY v7). Altho based on discord NSFW samples, its not going great, or even well..
@Mescalamba Even so, it seems that in reality it's the pony models and any future derivative are the most advanced models...at least in my mind.
@olternaut AuraFlow is definitely quite advanced, given Lumina 2.0 wasnt around back then, it was sorta sensible choice.