NOTE: Please read the version differences before you choose the model for your download.
This model is a merge mix of many different models (no merge weights greater than 15%), with a strong emphasis on semi-realistic characters - and of course tentacles.
Unfortunately there have been many good attempts at this by others - but none have satisfied my own tentacle fetish, so I decided to make my own.
And I have to admit as of right now I am satisfied with the results my build provides. I hope you are as well.
This model is weighed heavily on strong negative prompting (listing what you don't want) , and simple prompting on what you do want.
I would suggest using the standard VAE: vae-ft-mse-84000-ema-pruned, as well as using the newer UniPC sampler. Upscale results are good with RealESRGAN_x4plus at 4x - but may achieve even better results with a better trained up-scaler.
It is also very important that I note : all of the images I offered as examples had zero inpainting nor editing done to them. Straight out of the box results, which is what I ultimately was trying to achieve.
The negative prompt that I use frequently is:
(worst quality, low quality), extra digits, text, signature, bad anatomy, mutated hand, error, missing finger, cropped, worse quality, bad quality, lowres, floating limbs, bad hands, extra legs, extra arms, random body parts, anatomical nonsense, neko ears, missing arms, missing legs, extra ears, tongue, hat, elf ears, extra hands, hands with no body, floating hands, penis with no body, amputee legs, amputee arms, extra bellybutton, pubic hairs, puffy vagina, puffy pussy, badly drawn mouth, badly drawn eyes, badly drawn pussy, badly drawn nose, badly drawn lips, body out of frame, poor facial features, bad composition, mutated body parts, disfigured, error, jpeg artifacts, malformed limbs, fused fingers, too many fingers
Description
This is model version v2.0 FP32 full version of The Tentacle Project.
Two of the biggest difference between v1.0 and v2.0 is: 1) artifact cleanup from removal of several dirty models that I used in the original release, and 2) more emphasis on realism vs. semi-realism.
The greatest difficulty in making the model more realistic? Tentacles are not real (DOH).. And until AI advances some (or someone trains a model that outweighs this one) - this is about as best as it's going to get IMHO.
As a friendly reminder - if your rig can handle the larger FP32 unpruned model (this one), I highly highly suggest using it instead of the FP16 model. It is more textual friendly, and not as needy for negative prompting as the pruned version.
FAQ
Comments (23)
V2 really seems much more better. Nice 😎
Awesome! Glad youre enjoying it!
You are a genius
😁, thank you but honestly I'm just a tentacle lovin fool
How hard is it to make it more realistic while still being able to be less realistic / semi realistic? The samples provided use way too many 'airbrush' effect for mur taste.
For more realism, I would encourage you to download pictures of real people having sex with real tentacle monsters and train your own model. Id be one of the first to download it. If either the tentacles look plastic or cartoonish, or the people look plastic or cartoonish it will ruin the whole picture. There has to be balance. That is of course, unless you're into that sort of thing. 😆 But in all seriousness, it's extremely difficult because these two do not exist in the real world together (at least not in the way it's being used). AI is just not smart enough yet to generate real people + fictitious things and make them both look genuinely realistic in the same picture. BUT, that might change tomorrow, in which case I will make a v3.0 to accommodate those changes.
@brianclark079145 I would love to help achieve a more realistic version. But I have an integrated GPU and I can't get good access to any GPU collab services at the moment. Plus I know someone is already working on it so I definitely don't want to steal their 'thunder'. But otherwise I would be making it myself. I am looking into Loras and smaller models though and supposedly Loras can help achieve a more realistic look on any compatible model so there is that option that might do.
What Pos-Prompts i need for generating Tentacles?!
The full size model FP32 v2 or 1.0 non pruned version, are both text friendly for inputs. The smaller data models are good but require a little more work and inpainting to get great results.
There are no special keywords.
"Blond woman, pink tentacles f___ p_____" or "Blond woman, purple tentacles f___ a__" just as uncreative low weight examples. But you most definitely can get more creative with your prompts than that.
Just make sure to have a good negative prompt, the one I use is in the model cover page
Jell-O and Calamari
When the image is almost complete, it will look as if a sulfur filter has been applied, details are lost, quality decreases, the geometry of the face deteriorates. I use similar prompts as the users in the comments, but the image still looks weird. I kind of installed vae-ft-mse-84000-ema-pruned
same problem
I would suspect your upscaler or other filters being applied to the final picture (such as a custom lora, or script). If you have recently updated your ai base server (such as auto1111), it would also be prudent to check the most recent release notes to see if there were any changes made that could alter how the final picture is processed. In any event, it definitely sounds like a post-rendering process giving you those results. This model has no baked VAE, so it's not the model.
We can get content with using specific settings for all of our art generations, I know I am guilty of this. That sometimes a simple tick on or off in our settings can change everything.
There is also the possibility you are using an edited version of the vae rather than the original (less likely but possible, because there are edited versions with the same name out there) , in which case the original can be found here.
https://huggingface.co/stabilityai/sd-vae-ft-mse-original
Hope that helps
@johndoerandom33908 please see my note in original poster's issue
I really like your model, but I am having a problem with it. I am sure it is because I am inexperienced. I got a few really great tenacles in the early images, but as the images stacked up, they quickly went away.
Here is my prompt:
A pretty nude woman with long dark hair. A shiny black tenacle f______ p____. ((A black tenacle sucking on her n___le)). The background is a castle jail cell. extremely detailed hands, extremely detailed small breasts, extremely detailed face, extremely detailed face
My negative prompt:
monochrome, deformed teeth, deformed mouth, poorly drawn hands, watermarks
I am using INVOKEAI 2.3.5
Sampler:
ddim
Steps: 30
CFG scale: 10
Postprocessing
1: Face restoration (GFPGAN)
Strength:
0.75
I get a girl with long black hair.
I could really use some help.
Thanks
Hey there. Thanks for the interest. There are several problems with your prompt. First, tentacles is spelled wrong. If there are spelling issues in your prompts, the ai will "try" to interpret them, but most times.just drop that request and move on.
Next, is the "extremely detailed" segments. Might I suggest uhd, 4k, Fuji film, or ultra realistic as a single (or multiple) qualifier(s). The more specified requests that you add into the prompt - the further away you get from what you are actually looking for.
If you find that there are unfavorable things that are showing up in your generated pictures, use the negative prompt to eliminate them. For example, to get better hands, put "poorly drawn hands" or "deformed hands" or "missing fingers" etc in the negatives.
You should be careful with using () " " or {} because while they do put emphasis on the picture, they are seen as the most important emphasis potentially overriding others. Each AI program is a little different than the next, some like auto1111's provide more flexibility with emphasis qualifiers, while others do not. I am not familiar with invokeai so I cannot speak for it specifically, so I would consult the documentation or reach out to the forums to see how flexible those qualifiers actually are.
DDIM is a good sampler, but it is one of the more legacy samplers out there. For inpainting, and outpainting it's pretty darn good. But for prompt interpretation and generation, not so good. I would suggest using a newer more streamlined sampler such as one of the UniPC versions or possibly even a reworked hybrid of DDIM. UniPC samples at 20-25 steps far better than DDIM does at 40-50 steps, just as an example. So you're generating better pictures, more pictures, in less time.
Cfg scale - documentation suggests the higher the value, the more literal it tries to take your prompt, and this is true across all the generators. But in my experiences , it takes away from the AI being creative, damages the quality of the picture, and causes more artifacts in the final picture. 7-8 or any level in between will give you better results such as 7.3 or 7.7. It's a catch 22, because AI is trying it's best to be creative - yet the higher the cfg is, you're telling it not to be creative. Too low of a setting, and the AI goes into all out creative chaos.
Finally , you have to remember - it's AI. It's trying to interpret your negative prompts and positive prompts into something you want. This particular model is fashioned on strong negative prompting, and simple positive prompting. Each model is different in how it interprets the prompts. Which is why I left a detailed explanation of this model in the cover pages, and settings which will give better results.
I hope all of this helps you, and wish you the best of luck in playing with the model.
😁
@brianclark079145 Thank you so much for the detailed explanation. I will try the suggestions. Like I said I am very inexperienced and you have helped me so much.
Really an excellent piece of work!!!
Anyone have any tips on how to prompt better quality facial features like in the example photos? I'm getting good overall results with this model (amazing actually) but still struggling to get facial features that are more distinct.
This model is the best!
Dunno if you have seen this - older checkpoint, but has full training data available i believe... might be of use for a v3?
https://civitai.com/models/1169/sxd
pony?
It stands for “The TTP Project”.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.