This checkpoint, named Faking NSFW is my first attempt based on SDXL, an unlocked model designed to produce highly photo-realistic images. Built from a merged dataset, it shines with detailed prompts and excels at capturing realistic details, making it a solid tool for character designs or concept art. It is unlocked so it can handle NSFW content really well, one has no limits baked in with this model, but with great power comes great responsibility. Direct translation is king of the tree: fa (meaning tree in Hungarian) + king (Plain English) a play of words ;)
CFG, Steps, Sampler+Scheduler (after extensive testing):
CFG:
3.5 - 6.5
Steps:
35-40, 40+ for more details with diminishing returns
Sampler:
exp heun 2 x0 sde > euler ancestral > seeds 2 > seeds 3 > res multistep ancestral > dpm2 > dpm sde > sa solver > heun plus plus > dpmpp 2s ancestral > dpmpp 3m sde > unic pc bh2 > er sde > sa solver pece > dpmpp 2m sde heun > euler > heun > dpmpp 2m sde > unic pc
Scheduler:
gits > ays 30 > beta > beta 1.1 > karras > sgm uniform > simple > ddim uniform > exponential > ays 30 plus > normal
Actual results, any of these combinations can work (most times). The samplers in higher rows are more stable, schedulers from left to right go from most stable to less stable:
As you can see "exp heun 2 x0 sde" + "gits" combo is the most stable. "Unic PC" is highly unstable and only the most stable "gits" scheduler can help it, not worth using much. What is not listed was not tested or was so bad that no Sampler+Scheduler combo resulted in something usable on first few attempts, but feel free to experiment there are remote cases where for some odd reason they work even "Unic PC" with example "beta" can spit out some unique snowflakes :D
I usually use SDXL for the base idea (sometimes it's enough), then I go in with Flux at 0.4 to make things look more realistic Faking NSFW Flux - v5.0 | Flux Checkpoint | Civitai
IF you need to prompt for lookalike X or Y use v9 mostly, also v9 is best for prompts like polaroid, trending on Instagram or similar.
Description
Some corrections mostly face. A bit more creative in some scenarios.
FAQ
Comments (2)
Testing them all :)
v10 & v1.0 has the same filename!
So do I get v10 or v1.0 when downloading?
They get renamed on upload get them both rename if you need to filename on my disk is fine, not after upload unfortunately. Anyway version v1 is the baseline photorealistic version but has some minor errors in dark areas I would get v2 instead, similar more contrast and corrected errors.