Welcome to grabby realistic tentacles! This started out because I wanted realistic images of tentacle porn, and turned into a 13-month-long personal project to make a LoRA that could satisfy every tentacle kink I personally have. Hopefully it also satisfies some of yours! This LoRA (as of v5.0) is made from 7827 hand-labelled training images, with a lot of diversity.
Some recommendations, it seems to work just fine with most realistic checkpoints, but it also makes semirealistic checkpoints start to look real, and anime checkpoints to look semirealistic. For Pony, it seems to apply "double realism" especially to the face, which gets a lot of shadows, which I have represented in the showcase images. But, if you don't want that, and you just want it to "understand" some concepts, try dialing the strength way down, to like 0.2, because it seems that at a very low strength, it will still mostly obey your prompt, but not dramatically shift the realism.
NEW IN v4.1:
[x] Qwen-Image: By popular demand, well, two people, but they're probably popular people, Grabby Realistic Tentacles for Qwen is now available. See the note at the bottom about its limitations. I apologize if you don't have an RTX 5090 or higher. But this LoRA represents a key shift in the Grabby Realistic Tentacles universe. You'll note that the prompts have changed completely. The model will still make you things with old tag-style prompts, but it's best to be very verbose in your positive prompt if you want a detailed background. The model has been trained very differently to previous versions. Ideally, now you can just use normal human language to describe what you want. However, I'll still list new concepts that it now supports as "tags".
NEW tags in v4.1:
[x] "with suckers", "with suction cups", "with suction pads": Octopus-like tentacles! You can also use the word octopus, but it will spawn an octopus in your image most of the time.
[x] "<ribbed|rough|glossy|slimy> <COLOR> tentacles": The tentacles now have words for their appearance!
[x] "Phallic-tipped tentacles": They look like long dicks!
[x] "Masturbating": She is fingering herself.
[x] "Fingers touching her <vulva|clitoris|labia|labia majora|labia minora|vagina|thigh>": She is fingering herself specifically.
NEW IN v4:
[x] SD1.5: The model has been trained on SD1.5, it is nowhere near as strong as Pony/SDXL, but it does exist now for those on limited hardware. It's been trained on 512x512 imges, as well as 1024x1024 and 2048x2048 images, so that you can upscale things to HD if you like. It works reasonably well when you ask it to generate 1024x1024 images from scratch, but it's much better to first generate a 512x512, then upscale, and resample with a little noise. As with all SD1.5 stuff, hands and limbs are...problematic.
[x] Flux1D: This lora has also been trained on Flux 1.0 Dev. While not as obedient as Pony, due to the sterile and not-great-at-NSFW base model, it will make extremely realistic images. I'm hoping someday, someone with much more compute than me, can train a Pony-for-Flux version of Flux that understands how to draw NSFW images properly.
[?] Illustrious: This LoRA was also finetuned on illustmix v1.0, and seems modestly competent at following instructions. It's experimental right now, and I'm not happy with the cartoonish look it gets at a strength of 1.0, so I recommend a strength of 0.5 right now. I'm not sure what's up with my Illustrious config that would cause it to be cartoonish. But it is great at prompt following!
IMPROVED tags in v4:
The dataset tagging has been improved, to increase obedience with nipple related prompts. Previously the model was confused because isn't something that's touching a nipple, also touching a boob? The model would try to make a niptoucher that wasn't a boobtoucher. Now, if you include these extra tags, it will be much more obedient and reliable with the touching of boobs and nipples.
[x] boobtoucher: A tentacle touching a boob
[x] boobtoucher, niptoucher: A tentacle touching a nipple
[x] boobtoucher, niptoucher, nipsucker: A nipple sucking tentacle
[x] boobtoucher, nipsucker, petalsucker: A nipple sucker with flower-petal like sucker
[x] boobtoucher, nipsucker, cupsucker: A nipple sucker with cup-like sucker
NEW Supported tags IN v4:
[x] <COLOR> tentacles: The model now understands most colors, including odd ones like "olive", "dark wine colored", and "bergundy". It still leans towards reddish tentacles, if not explicitly prompted for a particular color.
[x] slimy tentacles: They are drippy and slippery.
[x] tentacle rubbing vagina: A bit hesitant to go all-in.
[x] niptoucher: A tentacle is touching the nipple, but not sucking.
[x] tentacle squeezing nipple: And is really going for it.
[x] suspended by tentacles: In the air! Might be supported from below.
[x] tentacle squeezing breast: encircled and bound!
NEW v4 Facial Expressions:
[x] blank stare: Put this in the negative prompt to have a semi-random non-blank expression
[x] <scared|piercing gaze|orgasm face|seductive gaze|sexually aroused|overjoyed smile|smiling face|feeling defiled|startled look>: How expressive!
[x] choked by a tentacle: Previously supported, but now, she looks concerned about it.
[x] brainwashed: Blushing, drooling, empty eyes, sweating, the lights are not all on upstairs.
v4 Artistic Flair:
[x] demon horns: For all your horny demon needs
[x] underboob: This shirt isn't long enough!
[x] wet hair: Previously all hair was dry always
[x] <melting|dripping|ripped|torn> <translucent|seethru|sheer|opaque> <COLOR> slime <CLOTHING>: Like "melting translucent green slime crop top", because I love slime clothes. Literally has nothing to do with tentacles but it makes the training process more fun for me.
[x] photorealistic/ultra realistic: If I didn't know any better, I'd think this was a photo!
[x] realistic: I can tell this isn't real, but I did have to look closely.
[x] semirealistic: The image isn't super realistic, maybe it's something about the eyes or the background, but it's not cartoonish.
[x] cartoonish: Nobody would be fooled at even a glance to suspect these are realistic.
[x] lactating: Seems to be much more reliable again
NEW IN v3:
[x] tentacle pit: The entire background is nothing but tentacles, she's in the Sarlacc
[x] catgirl: Ok let's give something to the furries
[x] slime hair: Hair literally made of slime, usually I put this in the negative prompt, I personally never want this, but you do you.
[x] <DESCRIPTION OF> tentacles: Starting to understand things like "leafy vine-like tentacles"
[x] restrained+hogtied: If you want your girl to be completely unable to plausibly escape, add hogtied (keep restrained)
[x] boobtoucher: The tentacle is touching a boob at all
[x] boobsqueezer: The tentacle is squeezing a boob
IMPROVED IN v3:
[x] Pony Support! We had this in v1, but v1 was trash!
[x] dry/sweaty/slimy/wet: Very much improved understanding of this, 'wet' in particular is a lot better now.
[x] tentacle in ass: Much better at this now
[x] tentacle in vagina: Much better at this now
[x] nipsucker, petalsucker: A nipple sucker with flower-petal like sucker
[x] nipsucker, cupsucker: A nipple sucker with cup-like sucker
[x] covered in slime: Much more reliable lubricant
[x] covered in cum: Much more reliable white lubricant
[x] cum: It comes out of tentacles
[x] slime clothes: Clothes made of slime! Try adding "translucent", "opaque", "melting", "ripped", like: "translucent green melting slime bodysuit"
Work in progress:
The LoRA currently doesn't understand a few things, and I'm working on it for V4. It still doesn't have the full variety of tentacles, like ovipositors, and milkers. It can generate penis-like tentacles, but I haven't found out how to prompt for it. I also want to have a way to prompt for an octopus style tentacle, a ribbed-for-her-pleasure style tentacle, a slime tentacle made of slime, and a slick tentacle that is smooth. I also want to add thick and thin tentacles.
==========
From v2:
[x] white tentacles
[x] black tentacles
[x] yellow tentacles
[x] slimy: Seems to be working now, not as reliable as "sweaty". If you want her to be slimy, use "sweaty", not sure why it's sweaty
[x] nipsucker: I learned that "nipple sucker" made the model want to render a nipple and a sucker, it was a bad training word. This is a tentacle sucking on a nip.
[x] petalsucker: The nipsucker has flower-petal-like grabbers.
[x] cupsucker: The nipsucker has a suction-cup-like grabber.
[x] tentatug: The tentacle is tugging on the breast, usually by the nipple
[x] tentasqueeze: The tentacle is squeezing so hard, there's visible skindentation!
[x] offsucker: The tentacle missed the nipple. It's sucking something, but its aim is off. I usually put this in the negative prompt.
NEWLY BROKEN IN v2:
[ ] lactating: She used to leak!
=========
From v1:
[x] tentacle in vagina: Get in there!
[x] tentacle in ass: Plug!
[x] tentacle on breast: The tentacle is laying on or across the breast
[x] restrained: held down, or up, or sideways, but held firmly
[x] choked by a tentacle: Currently more like a tentacle choker, the women don't seem very concerned
[x] naked
[x] translucent slime clothes: Clothes made of slime! It's not a great material, it's completely see through!
[x] opaque slime clothes: Thicker globs of slime for modesty.
[x] melting slime clothes: It's too hot in here, my clothes are melting!
[x] ripped slime clothes: Not sure how one rips slime, but the model can do it!
[x] covered in slime: Very slimy
[x] covered in cum: Absolutely slathered
[x] cum: It's somewhere in the image
[x] cum in pussy: Something must have JUST cum in there
[x] dripping pussy: Usually with cum
[x] dry: Her skin is soft and dry
[x] wet: She is shiny and dripping, with a clear substance, might be slime, probably is water
[x] tiny breasts: Flat chested
[x] small breasts: Slightly smaller than average, but nice and perky
[x] medium breasts: My wife has these, and so do most women
[x] large breasts: My wife wants these, and so do most women
[x] huge breasts: This is utterly impractical, these are ridiculous, unrealistic knockers, you need medical help to walk around with these gazongas.
[x] red tentacles
[x] orange tentacles
[x] green tentacles
[x] blue tentacles
[x] purple tentacles
[x] pink tentacles
[x] brown tentacles
The model also slightly understands these (I use these in the NEGATIVE prompt because I don't like them):
[x] medusa hair: Her hair is turning into tentacles!
[x] slug: There is a floating, disembodied slug, it's not a tentacle.
[x] tentacle horns: Like demon horns, made of tentacles
[x] tentacle ears: Like cat ears, made of tentacles
QWEN v4.1 downsides:
It's not exactly where I would like it to be, but it exists at all. It has a handful of major downsides, first, it's a BF16 lora, which means that if you have a quantized model, it will use an insanity of VRAM compared to your quant. I haven't tried making it smaller yet, but as a side effect, it may run you out of VRAM even on very nice GPUs. This is especially true if you use nunchaku quants that drop it down to 4 bits. This LoRA will expand the 4 bit quant back up to 16 bits, and it'll take up 4x as much VRAM.
The LoRA is free to use for personal or commercial use. You don't need to credit me. However, if you figure out how to fix any of the current bugs, you must tell me in the comments so I can fix it for the next version! If the model isn't doing something and you want it to do that, let me know! Due to the lack of real photos of this, this model has exclusively been trained on images that I personally generated.
Description
Version 4, many more sample images, especially of nipsuckers, and more expressive faces
FAQ
Comments (7)
Oh mighty Turlyag, would you consider an IL release for this LoRA? I used it extensively when I was using Pony, but the adherence with IL is so good I can't see myself switching back.
Haha sure, what's your favorite checkpoint. I'm training v4 now, and adding SD1.5 and Flux1D support, so it's a good day to add IL. What's your favorite realistic IL checkpoint?
@turiyag You're amazing. Honestly, ILustMix is my go-to, not because it is super realistic in and of itself but because with proper LoRA support it is so insanely flexible and can be realistic when you want it to. I've also had great luck with compatibility when training LoRA's on it. Would that be a good option?
@Shiggles You sadly only get to know if a checkpoint is good to train on, after you're all done training on it. But I'll try it for iLustMix! To get really technical, a LoRA is a modification to a given checkpoint, which should work on "similar" checkpoints. But it represents the difference between that checkpoint and the training dataset. So if I train a bunch of images against the basic Stable Diffusion 1.5 checkpoint, the original, then the LoRA represents not just the training of tentacle based stuff, but also the difference between an average SD1.5 image and realism. So it starts to encode things like "don't look like anime", and "don't make an image without a woman in it" into the LoRA. This "takes up space" in the LoRA's head, if you will. It's like having to explain Shakespeare to an English speaker, vs trying to explain Shakespeare to someone who only speaks French. You have to also teach them English for them to understand Shakespeare, which is a much harder task.
But if you train on a checkpoint that's already "close" to what you want, like, it produces XXX images, it understands naked people, it understands tentacles, etc, then the LoRA has the ability to encode the differences there. So for example it gets a lot more "space" to train concepts that are missing in the source checkpoint, like a tentacle sucking on a nipple.
You can see with the v4.0_Flux1D version of this LoRA, for example, that a lot of the LoRA is "taken up" by just having it draw NSFW things, and drawing tentacles at all. It can't handle the nuanced things about the LoRA in the same way that the Pony checkpoints can, because Pony was explicitly trained on NSFW content and art involving tentacles.
This is most visible in the dark eye sockets in the v3.0_Pony/v2.0 (SDXL) LoRA models, because those were both trained on the base checkpoints of PonyXL6 and SDXL1.0, so they encodes the difference between an average image and realism. But then when you use a realistic checkpoint, then it's "double realistic" which unfortunately doesn't mean twice as realistic, but instead, since Anime eyes don't have eye socket shadows, but realistic images do, it adds extra eye socket shadow, but the realistic checkpoint added eye socket shadow in the first place, so instead we get extra shadowed eye sockets.
The v4 LoRA is going to be trained against realistic checkpoints, so hopefully won't have this issue.
So, long story short, I've got another 8h left on the training run for Pony, and once that's finished, I'll test it to make sure it's working, and if it is working, then I'll post the v4.0_Pony LoRA, and start training on iLustMix.
@turiyag Ok, first of all, thank you SO MUCH FOR THE INFO DUMP. I really appreciate it! It's definitely going to change my LoRA selection from now on when training, I've always just defaulted to whatever checkpoint is my favorite at that moment.
Also, thank you tons for trying the training on ILustMix. I can also do some research into ideal alternatives that fit a more realistic style.
@turiyag I can't believe I didn't think of it before, but AlchemistMix V4 seems like a prime candidate for a more realistic IL model!
@Shiggles The illustrious version is out!