Pandora’s Box (NSFW)

Pandora’s Box: NSFW Nude Pinups for SDXL

“They gave Pandora a box. Prometheus begged her not to open it. She opened it. Every evil to which human flesh is heir came out of it. The last thing to come out of the box was hope. It flew away.” ― Kurt Vonnegut, Timequake

After a few weeks of merging & testing, I am pleased to present you with Pandora’s Box. This model has exceeded my modest expectations, quickly becoming my favorite SDXL checkpoint for generating classy full nude female pinup photos (and much more). Good at photorealism, 2.5D & wide variety of illustration styles. It plays well with most of the LoRAs that were tested. Some LoRAs even boosted or changed the photo-realism quite nicely. It seems quite capable at standard non-nude prompting as well, achieving different but often equally interesting quality results when compared to most of the popular SDXL models on this site (as far as I can tell). I find it has particularly playful tendencies, perhaps in part due to the anime & furry nsfw models that I merged in earlier on, giving it a cute popart cosplay edge.

V1: Puss & Boots:

I’m still getting know this model, and SDXL in general, but here are some things I’ve noticed so far. Please let me know if you find any alternate winning combos.

Results depend greatly on CFG (5-10) and steps (20-60+). I usually start near 6 CFG & 38 Steps, to get the ball rolling. Euler A or DPM++ 2M SDE Karras.

As is common with SDXL, sometimes hands & eyes can be off, depending on +/- prompt, sampler, dimensions, CFG, Steps, hi-res fix, etc. I recommend copying the basic generation data parameters from the example images

OR better yet, set up an XYZ script to see effects of subtle changes in steps & CFG on a given prompt & seed. (sometimes just a few steps & 0.2 CFG can make all the difference)

Hi-res fix parameters: (hi-res steps 8-20 / denoise 0.2-0.5)
Hi-Res Fix is not always needed, but it can help refine various details like eyes and fingers. Sometimes it goes too far & makes it worse, or removes detail from the skin texture, if the equation is not right. In these cases I might use Adetailer for additional face & hand perfection. No need for Refiner (I’ve been meaning to test it more, but initial tests weren’t so hot).

Also keep in mind, that the latest version of A111111111111111 introduced the ability to load a separate checkpoint for Hi-res fix. Allowing you to generate a complex composition with SDXL, then run it into your favorite SD 1.5 checkpoint for the hi-res fix pass, to get the added detail you want. Hi-res fix even has it’s own +/- prompt boxes now! Nifty Biz!!

Specialities:

Full frontal female nudity / pinups. (Not always perfect, but a fair bit more consistent than some of the earlier NSFW models I tried)
Good at Cosplay goth fashions, group photos, detailed interiors, text logos
Plays very well most LoRA’s I’ve used (photo style, art style, NSFW)

Issues:

No proper full nude men with penises (men have blank plastic Ken-parts)
Can suffer from plastic skin & mutations (much like most XL models)
It’s not good at pubic hair
No proper assholes as of yet, butt I’ll try to rectumfy this, along with the other issues.

Merge Procedure:

This mix was made mostly with Supermerger, using a concoction of several models & additional loRAs. The procedure was very experimental, but I like to think there was some method to the madness, landing somewhere between trial + error + instinct. After many merges, the most interesting results came when I started back-crossing newer versions of models that were used earlier in the mix, using the “Train Difference” setting. (much like back-crossing plant-strains, to strengthen the genes & phenotype). For example, zavychromaxl_b2 was used at an earlier stage in this mix, so when zavychromaxl_v10 was released, I decided to use it as the A model, with my unstable NSFW beta Pandora mix as B model – minus the older (previously merged) zavychromaxl_b2 as C model, all using the “add difference” & “train difference” settings. This kept some of the sexy themes I merged previously, while introducing more stabilization and versatility in the direction I was going for. Here’s the equation for the final step in this merge:

zavychromaxl_v10 + (Pandoras_Box_BETA – zavychromaxl_b2) x 0.401_trainDifference

= Pandora’s Box 1.1

List of merged models:

Please let me know if I’m stepping on anyones toes, or if you wish to be unlisted.

Some of these were only used in microscopic amounts. I never expected to share this when I started it, but I’ve been having fun with it & wanted to give back to the community, in case anyone else finds it useful. Bless you all.

Massive appreciation to all the trainers & mergers out there !! CivitAI !! Developers, artists & wizardly nerds !! The lineage of knowledge, skills, craft and style that goes into this strange new medium is epic !!