Realism by Stable Yogi Pony V6.5

Seems to work fine for me.

Without seeing any examples of what you're trying to do, I can only guess, but maybe you're using LoRA, and/or prompts that conflict with your goal of producing an anal sex scene. You may also be trying to use the wrong prompt, as for instance could be describing it as "anal cowgirl sex" when you should have used "anal cowboy sex" instead.

Share us a better checkpoint then if this is not good

I might agree to a degree. I love it, but it still has a few "pony quarks" that ultimately leave me wanting something else. What that is, IDK. Base XL is too much work to lol, NoobAi MIGHT be onto something, but eh. Even then there are still too many holes in what all the "at home" stuff is capable of. This one makes sense at least

@jumpforjoy I doubt it's about "not being good" as much as "Pony has it's perks but not much else"

@D3monCeps what do you mean with "still too many holes in what all the "at home" stuff" which home model do you refer to?

@pink0909  Many! Some basic actions and poses and such are lost concepts to most models, from SD1x, top Pony, to many others here. Most seem to be either "portrait studio 1 million" "AWA generator 10 quadrillion bajilliion" or some odd mix of the 2. We shouldn't be forced into controlnets for a MAYBE just to attempt achieving pretty normal and mundane human poses and action, yet this "far" along, that's exactly how it still is. For example armwrestling, tug-of-war, and don't your dare try driving. The mega corps and all their censoring don't even flinch when you ask for some action. this stuff... lol.

@D3monCeps megacorps are lame thats why local generation is king, i was suprised that on huggingface a huge community is uncensoring LLMs (the ones Megacoprs build in anyoing supernanny "guidelines") the best ones are called abliterated. Real good. Only Downside you need a Gpu with 10-12gb Vram at least (8gb and 4gb can work but with significant drawback ind Modelsize and speed)

@pink0909 The nanny stuff is awful. Sadly, they can still do stuff that the local stuff simply can't (or requires a blood and silicon sacrifice to every deity in existence to give you something the MIGHT be inpaint-able) You'd think with all these models here, one of these geniuses would have figured out how to replicate what the big leagues have done, minus the sensor-fu. I'd like to think I just haven't run into the "right" models for my whims, but we'll see. Will have training capabilities eventually to attempt plugging the holes (probably wont, because if it aint NSFW for the sake of it, just lol

@D3monCeps well it depends how much you ask for, i am a seasoned guy starting with 368-Pcs and win3.11 with Paint xD. Then i basicly bought every PC generation (build my self since prebuild PCs are overpriced). I hopped between AMD and Intel depending who had the price/performance advantage in that year. Back in the day Selecting stuff for "inpainting" was a real hussle, you basicly had to become a "photoshop" master just to make semi real fakes (or work with Photoshop in general). These days We have now Segment-Anything, and even automatic lighting correction, Depth-Estimation,Pose estimation, Logogeneration, Video and more - all with a click of a button (Each one of them is a Huge Step Forward). I remember a >25 year old meme(called Demotivals back then) image joking a programm could atomaticly make nudes with the MagicWandTool (select all tool). I am really happy since i get high tech (Stable Diffusion) for old Foto restauration (analog + digital), which is close to magic since it goes beyond restauration it can redraw a crappy QVGA-Image into HighRes-Quality (sure you need a Few click more than one xD but heck its possible!). So maybe changing your POV might make you happier :) .

@Stable_Yogi thanks. never knew about that section. i'm used to the info being under the preview images.

someone already pasted the answer to that in this comment section, son I am just going to quote

"You need quite a bit of negative:

blurry, blur, motion blur, depth of field, shallow depth of field, bokeh, glow, soft background, out of focus, soft focus,"

@pleaserphotography236 That's literally my comment.

@Electroverted that´s why is quoted word by word in quotation marks and specified that it is a quote from someone, and not mine :)

@Stable_Yogi 
Thanks for responding, but I wasn't asking about me, I was asking about you.

For me, I do basically all my AI related stuff on my own computer, so as the username I have suggests, I'm here to get models, LoRA, and such for me to use. These changes directly don't have much impact on me as a result, although if they were to really screw with what you do, and why, then people like you may stop making the models, LoRA, and such we use.

That is what I was asking, it's good to hear you at least currently intend to keep at it, although what is the impact on you, are you doing this for fun, so whatever, or is this a means to make money, so would it really impact that for you by your estimate?

where can we find models, lora that they censor on here now?

V3, means version three, V5, V4, and so on mean the same. Yes, this means an old version, V3 is still very popular, I like V6 more, but try them yourself if you're doing this on your own computer, and use what you like. Some versions of a model, and some models will do some things, better than others, so you may end up using a collection of them, for different types of things.

VAE indicates it has a built in VAE, which is fairly common with various models. VAE stands for variational auto encoder, feel free to look up info on them, but basically you would want to be using one when generating images.

Stable Yogi started putting info on what the other parts are more clearly in the "about" section in the column to the right, under where you would find the download link, as well as other info, so you should look that over. Although:
FP16: Basic version.
BF16: For use with higher end graphics cards.
DMD2: The model was merged with a hyper, lightning, or whatever you want to call it LoRA, this lets you generate images with less steps, and therefore faster, but you lose some quality as a result.

@AImodels4me2use 

i was totally confused with all the v3_vae fp16 bf16 dmd2 stuff and now i finally get it lol especially the fp16/bf16 thing, you saved me dude, seriously thanks a ton!!! 🙏🙏

Adding maybe useful addition:

V3: Version 3, increasing version is newer, now V6

VAE: Image encoder and decoder, needed for image read => i2i or i2v image gets encoded to latent. Needed for image write => t2i and i2i, or t2v and i2v, when decoding from latent to image or video.

"VAE baked" means the VAE was embedded within the model, so if the baked comment is missing, you need an additional VAE (sdxl_vae). Without VAE you will get black images, with wrong applied external VAE similar. Care for settings.

fp16, bf16, fp32, fp8 describing "float point in bit" accuracy. fp8 < fp16 < fp32 regarding versatility, quality, precision. The higher the bit count, the more compute power (GPU/VRAM) and inference time is needed. The bf16 model though is a different breed, mostly used for training, the least "tamed model version". Similar quality as fp32.

Rule of thumb: fp32 => 24BG VRAM, and fp16 => 16GB VRAM, fp8 => 8BG VRAM (highly depending on model and skill).

DMD, lightning very compressed model versions with specialized schedulers, fewer inference steps, but less quality and variety in "single model only" use. Also used for inpainting or multistep upscaling, for increasing(!) quality of a given image.

Q6, Q8 if you encounter those models this means "quantized in bit". Those are "reduced" models, resulting in lesser load on VRAM and GPU. Mostly comes along in GGUF not safetensor format. The smaller the quantization Q2 < Q4 < Q6 < Q8, the lesser accuracy, the more error from original fp32 model. Q8 model has ~ <2% error rate ("pretty good"), Q6 ~ <10% error, and so on.

DMD, fp8 and Q models are mostly needed for smaller consumer rigs (laptop/low vram), POC prototyping, for new concept, upcale and inpaint - when there is no need for high precision (fp16 / fp32). fp8 models are coming strong, but only new gen stuff end 2025 beginning 2026 (due to cuda design decisions. ngl though, f*** NGreedia, f*** OpenAI).