Vēritās = truth, verity, reality, true nature, real life
Note: Preview images are generated without the use of LORAs
NoobAI ver. [v1.0]
About
Originally this was my attempt at creating a photo-realistic model with the prompting capabilities of NoobAI or Illustrious models, but NAI/ILXL isn't quite there yet when it comes to photorealism. This model can do a bit of realism but it is nowhere near to what I would consider as photo-realistic, which is why this model is labeled as Veritas instead of Fidelis. I also tried to reduced the 'realism' to make it more 2.5D rather than 3D but it still turned out too realistic, oh well..
Tips and Usage
Since I used my HomoSimile XL model as the base, it started out very 2D focused. And the transition from 2D to 3D is not as straightforward as I had hoped. So there are still some kinks I need to work out. I highly recommend using ADetailer and Hires Upscaling. I don't know if it's a common thing for ILXL/NAI realistic models or just mine, but sometimes certain features turned out low quality, especially the faces and hands.
Style tags obviously won't work as well, but character tags do still work. But whether those 2D characters look good in 3D depends on the characters themselves. Some characters do look pretty good, but others look a bit awkward, and some characters don't look anything like their 2D version at all. You can try using character LORAs, but I haven't tested any yet..
My Settings
Sampling Method: - Euler a (less detailed and smoother textures)
- DPM++ 2M (more detailed lines and textures)
- DPM++ 3M SDE (more detailed lines and textures)
Sampling Steps: 25 - 35
Upscaler: 4x_NMKD-Siax_200k / Remacri
CFG Scale: 3 - 5
Clip Skip: 2
VAE : Already built-inPositive Prompt:
realistic, photorealistic, very awa, masterpiece, best quality, absurdres, perfect quality, detailed background, intricate detailsNegative Prompt:
lowres, worst quality, bad quality, bad anatomy, signature, username, logo, watermark, artist name, bad hands⚠️DISCLAIMER⚠️
Please do NOT generate illegal content with this model. You are solely responsible for any unethical use of this model(s)
Model License
Description
v1
What's new:
Initial version
FAQ
Comments (60)
oh my gosh!~~ can't wait to test.
appreciate you! ❤️❤️❤️
Goodluck! I hope you like it.. :)
Better to use, more obedient, and more expressive. But the texture is still like a figure model, hopefully to evolve a more realistic version
______________
cant wait to try.
Just to clarify, this model is meant to be 2.5D and semi-realistic, not photorealistic. I might make an update for HomoFidelis which is meant for photorealistic. But right now ILXL and NAI models aren't quite there yet when it comes to photorealism. I tried making it very realistic but it also reduced the prompt obedience.. 🤷🏼♂️
Let's GO!!!!!!!
Oh, I'm going to be SO obsessed with this.
so, what the diffrent between this and Homofidelis pony V2?
1. Better prompt adherence
2. NoobAI-based instead of Pony so you can use a lot of character tags
3. Definitely easier to prompt, no need for score_1 to score_9
4. Can do natural language as well (although tags are still preferred)
5. HomoVeritas is mean to be 2.5D or semi-realistic while Fidelis is meant to be photorealistic.
6. Can handle emphasis, for example (huge penis:1.2), way better than Pony models can
@Kairen92 Thank you so much ❤️
I'm a noob to NAI and used to Pony tagging with some NLP thrown in... I see in your examples a lot of nonsensical tags (very awa, absurdres,etc)... do you know of a good quick ref cheat sheet /Noobs for NoobAI guide out there by chance?
@jtmichels NAI is just a 'finetuned' version of illustrious XL. Think of it as an 'enhanced' version of ILXL since base ILXL only uses danbooru dataset but NAI uses both danbooru and e621. The only NAI exclusive tag that I used in my examples is 'very awa' which is suppose to be a quality tag similar to 'score_9, score_8, etc' for pony. Although the difference is mostly unnoticeable, I just included it because why not. 'absurdres' is an actual tag used by danbooru btw.
Personally I find it very hard to go back to pony after using IL/NAI. It's just better in a lot of ways. Pony was good for it's time but now it is kinda obsolete. And if you're thinking of moving to IL/NAI, just do it. It's prompting is pretty similar to pony and it can handle natural language way better than pony can..
You're a legend
What are the differences between this model and Homosimile V3.0 (Illustrious)?
Which one would you recommend to use?
This one is for 2.5D or semi-realistic images, Simile is for 2D.
If you like 3D and realistic-ish style then go with this one, if you like 2D and anime artstyle then go with Simile..
Thank you!! I've been so excited for this.
I'm a fan of all your models, but getting a bit confused on differentiating between them and best usecase scenarios. I train LoRAs on your Pony Simile with photos and it creates a pretty nice realism but able to be stylized very easily with prompting. When I train on Pony Fidelis, I cannot get likeness to compare to original subject nearly as well.
But that's wasn't the point of my comment! lol
Could you explain a little more the models/naming convention? Simile (comparison using like or as) is clearly the most stylized. Veritas (truth) and Fidelis (true to the original) are what trip me up more. Which is the true truth, forsooth?
Lmao sorry, I know it can be a bit confusing for people who aren't familiar with my models. I used those names back during my SD1.5 model days because I'm not good at coming up with names and now I'm kinda stuck with it lol..
But basically it's like this:
Fidelis : Photorealistic 3D, real photos, etc.
Veritas : 2.5D, semi-realistic, etc.
Simile : 2D, anime, illustration, etc.
(there used to be more but right now I'm just sticking with these three)
Btw, I wouldn't recommend training on my models as it is not designed for that. If you want to train loras for pony models, just train on the original pony diffusion xl model..
Thanks for the insights on the model naming! However, LoRAs are much more effective to maintain the finer grained details if they are trained on the model they are used on. Training on vanilla Pony then using on Simile would map the weights XYZ to embeddings that have since been finetuned / moved /merged /shifted which can cause artifacts, hallucinations, deformation, etc depending how drastic the change in weights are from the original. (Style and Concept LoRAs are much more flexible though.) Also, if I was going for realism and likeness, vanilla Pony would net me pretty bad results.
Sometimes highly merged models aren't suitable for finetuning because the gradients have become really erratic and noisy due to all the different models and more fragile, susceptible to burn/ exploding gradients/getting caught in local minima/etc... but I see no reason why your models wouldn't be suitable for training LoRA on, but hey there's something new to learn with AI every day! Maybe I'm missing something
omgoodness, this is perfection.
Ty so much!!
I hope you like it! :)
Almost perfect, but if there are any way to create better big nipples?
Try tags like "puffy nipples, large areolae, long nipples" and maybe use emphasis on it like (large areolae:1.2)
@Kairen92 Most time it works well but sometimes they are just like two tiny meatballs, not so good as homosimile.
@jsjem if you dont want the puffiness and just want to make them big just use "large areolae"..
Thank you 🙏. Also, I've been running into the roadblock you mention regarding character faces. Countless iterations of Bolin (LoK) and haven't generated a passable face. Worse because the rest of the images have developed into a 2.5D style worth extracting.
Hi there. Are you using lora or just character tags because Bolin doesn't have enough data on it's training images to be able to use character tags consistently. For characters with low datasets it's better to just use loras..
@Kairen92 That’s where things break down. The styles of the LoRAs I’ve used to generate Bolin are all too cartoonish and 2D to not break down as soon as realism comes into play. And I haven’t worked out how to correct for those traits in my prompts with any real consistency
@TsarDicholas Ah yes that can happen, especially if the lora is too 2D focused, so its not really on the checkpoint but rather the lora itself. You can try lowering the weight of the lora but it also makes the character look less like how they actually look like. Can you give me the link to the bolin lora?
The Tensor.art version now it is tilded as "duplicate", and the duplicate one it lead to a undefined version. Not even a 404.
Urghh, dammit. I think something got messed up when I tried importing it directly from Civit. I'll fix it later, thanks for letting me know..
@KubutHado I already reuploaded the model and tested it on the generator. Everything seems fine but do let me know if its working on your end. Thanks!
Kairen92 Still treat it as undefined. I'm following the link that show up telling me that is a duplicate from the one that upload it, this is the link to the "duplicate" https://tensor.art/models/889872154760378254 the original one: https://tensor.art/models/889872154760378254 if you can, tested with a second account, that could give some answers. And also, thank you for take your time to test it, because it did work quite well when it was around.
KubutHado Both of those links that you gave me are the same thing though?
I tried testing it with a new account but it doesn't show up. Probably because of the NSFW stuff. Tensor is updating their policy so maybe they're blocking my model somehow?
What about my other models? Do they work?
Kairen92 I just tested, and the others are 404 for the change of the rules, but they do still work if one had them bookmarked. This is the Veritas I have bookmarked https://tensor.art/models/883715075402118087 and as you can see, it send the duplicate message. On your account, neither of the models are show for the change of the rules.
KubutHado Hmmm, the one that you bookmarked isn't uploaded by me, but they did credited me. For now, let's just wait until Tensor is done with their policy update or whatever. After that, if it still doesn't work, I'll either have to reupload them as a new project entirely or just change the cover and example images.. Urghh, god I hate tensor.. 😩
Kairen92 It is rather sad that tensor, which can generate better (face) images than here, had to go in the way it goes for the faults of the credits cards, but the page is know for being bugged in general. Still, it did allow me to generate a lot of arts with the Veritas, and at the very least I can download it here if for the future they will only be possible to generate image on your own PC for how things are going: bugged pages, bad generate pages, or very stupid laws/rules.
Kairen92 I do apologize for mingle on this conversation, but I did indeed come here asking the same our fellow user question. And I do believe the real issues is that you may have tagged wrong the model. If you can see it on your account, it does not mean everyone can see it.
It is either change the tags, focus or any description toward NSFW or simple erase it, making the reprinted to worked again. Specially because the reprinted model is not banned on Tensor. I do apologize if this does sound a bit insulting, but that is the only way I can see, for the ones who use Tensor, an away for using your amazing model.
Because here on Civitai, the models work yes, but for how the page is, it gives a thousand of restriction when it comes to models and even the results are not the best in terms of details, as Tensor, still with nsfw ban, still do generate far better results.
OMG, just been away from XL for 3 weeks, and we have such a big gift? Thank you my friend
Hope you like it.. :)
Kairen92 I've tried for 2 days, and single character result turns out pretty nice. And for 2 persons interaction can be a bit challenging. It really depends on the prompt or Lora a lot. Maybe it's because I'm a bit spoiled by the Chroma (FLUX) model recently, it really makes the prompt engineering less important, so the user can try on more imagination.
saka83838383 Oh you can't really compare Flux or Chroma with NAI or ILXL, they are totally different. While NAI does support some natural language prompting, it still heavily rely on tags.
You can always try a mix of tag and captions. and then lower the CFG scale. It's the easiest way to get creative results..
Also, try to avoid using unnecessary loras, I've had some people using 4-5 loras at the same time and setting the weights to 1 and then DMed me why then can't get good results..
Kairen92 That's very true. I apologize for mixing that up. They are two different prompting styles. I would try more later on for sure. And additionally, I also found the facedetailer is quite critical to have a nicer result, isn't it?
saka83838383 Hmmm yes unfortunately. I think its because NAI is suppose to be a 2D model so when transitioned to 3D, its not good with human faces (furry faces are fine strangely). I even tested other NAI realistic models and the same thing happened too. So I highly recommend using face detailers..
Kairen92 I see, but considering the super fast generation speed it can do, (15 seconds with facedetailer compared to 70 seconds in Flux) Facedetailer is not a big deal, And btw, I just found a simple prompt structure, good at composition of multiple characters. Would share some result later.
saka83838383 Cant wait!! And good luck with the gens lol..
In love with the 2.5D style, and this model of yours as well as all your other ones are definitely the best out there. I am sad that I have a very specific style I like and have been able to use your Veritas SD1.5 model plus some loras to generate it (still requires some workaround but I managed), but those loras do not work with this XL model. However, I am happy for everyone around here that you continue to create such awesome models and that they won't stop getting better!
Thank you so much. I hope you'll find a way to recreate that specific style with this model as well.. ☺️
Hi there, did you hide the model on TensorArt when you uploaded it? I cannot see it anymore! Thanks!
Hi there, I did not. But someone uploaded my model there without my permission so Tensor thinks mine is a duplicate or something. I might need to reupload them again..
Kairen92 I think you can claim it via TensorArt's discord channel...
GloberTrotzer Probably, but I'm gonna try deleting everything and then reuploading it again. I think it's because I tried importing it directly from Civit the first time..
Great in solo and solo focus, but always out of control in over 2 guys and anal sex
Does anyone have a good alternative to this? I bid on it hope it gets back onsite.
I am a fan of yours, honestly I really like the model versions you have created, from semi-realistic to hyper-realistic, I am just curious if you plan to update this version, it really creates beautiful images.
That would be amazing ! :D
Hell yeah !!! :D
Wolph!! ❤️
@Kairen92 Hii !! :B
Please, excuse me, could you tell me where to find a WF with the ADetailer or Hi-res fix you use with your merges to get such flawless faces and glans ? Thanks in advance
I'm sorry I mostly use A1111 for my gens so I don't have the WF for it. But you can use any WF that has Adetailer on it and it'll probably work just fine. And as for the Hires Fix, I used Remacri which you can find on this site. Good luck!
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.