Hallucinations
why?
It's old technology that used to do that because the parameters used for it weren't enough to make it learn about things like human anatomy, so people resorted to fixing such problems by fixes like "inpaint" (you mask the problematic region and have the model redraw it), "control nets" (you have an existing image that isn't problematic and use them as a structure for the new one), Image to Image (a model without those problems is used to paint again the picture accurately), etc.
People like me continue to use the old technology because we're not convinced by the main alternatives like Stable Diffusion XL (which doesn't have a training as robust, so all its finetunes and loras look kind of similar to one another, like a vanilla flavor you can put honey on or chocolate chips but it's still vanilla. For his next version of https://civitai.com/models/139565/realistic-stock-photo , PromptSharingSamaritan had to go back to this technology because XL could never be as detailed), Pony (still SDXL technology that became popular only because of its performance for Not Safe For Work material), or Flux (Flux is great when you have something in mind and are willing to use entire paragraphs explaining what you want to it - BUT if you want to see a cool picture with some concept, Flux will plainly give a boring image, this old technology will be very creative about it and will add details and objects you didn't ask for. For a prompt like "girl in super metroid", this tech will add a lot of what you'd expect to see, Flux will output something simple and boring.)
Here's flux on the left and Chip & Dall-e on the right. I'm not saying Flux's output is bad, or worse, I'm saying on the creative department it'd never add what C&D adds to the image because I didn't prompt for it.
Another reason is people moved on to other technologies and thus nobody fixed the problems of this one, so newer model versions using it at the very least will require you to produce many pictures until one is good enough.
Also, it's funny how even Flux produced a hand with too many fingers, and that's state of the art.
I totally think like
@Yntec
said, old technology consists mainly with images with little prompt to nothing and produce an amazing sense of creativity.
SDXL and Flux are just the best with anatomy (while sometimes using inPain ControlNet plugin).
I prefer without any waiting time using Stable Diffusion 1.5 which is more surprising and elevate my visualization of creativity. Each Stable Diffusion 1.5 (with nice models such as Incredible World series or Gacha (my favourite)) produce a treats and sweets for my own visual pleasure and with prompting not fully what it adds! That's that which defines arts and creativity!
PS : I like even how fails produces amazing surrealistic visual amusement 🤭!
I really appreciate such feedback FashionStash! My favorite is Cryptids, it has full coverage of subjects I want, the most adorable chibi girls I've seen, a whole zoo of animals ready to be generated surpassing furry models, outstanding landscapes and backgrounds, incredible creativity and compositions, and it does objects better than models designed specifically to make objects!
rpgicondiff picture of treasure chest
I guess its name was unfortunate, since its inception was to create animal fusions called cryptids, all its potential remains hidden, and, of course, its complete lack of ability to make photo realism.
It's great to be able to talk about this passion of mine, it gets more niche by the day and talking about Stable Diffusion 1.5 tech feels like talking about VHS cassettes, even though it's very recent.
Of course "Lady Nostalgia" is making our hearts pulsing when we're remembering the old days past to discovering AI Open-source for generating stunning wonderful arts with our Stable Diffusion 1.5, haha, exactly, like our old beloved VHS or even walkman with K-7 instead of nowadays MP3 players.
I'm in a community in Discord, a localized one where people pushes in front of the scenes the merits of Flux... Or Stable Diffusion XL, but theses peoples are always in the need to use plenty of variouses external plug-ins to achieve their creative generation.
I'm not saying that Flux or SDXL are not great enough to get prompt to arts and so.. But I'm saying that SD1. 5 can really be a powerful standalone technology, I'm not the only one in the world saying that, and I'm trying to often point it to new person that want explore AI images to think about their arts instead installing plenty of high-volume diskspace consumption, and then instead using the hyped new technology, starting by the root of the beginning and see what was really the AI strongest power : creativity. No more, no less!
Cryptids, you said Yntec, I will try it then, I also have a great preference in surprises than having exactly what I prompted for!
I thank you by the way to port plenty of fantasticulous models at HuggingFace, you have a nice taste when it comes to the creatives models. And I'm never going to stop following your works, here at HuggingFace. Always a new sweets to test and love regularly!
I love also when you bake some models together to create amazing new way of creative models (as the last Memoji Remix).
I use many of my AI generated images to make TCG cards, just for my own rolist tabletop games pleasure (such as Magic the Gathering, Yu-Gi-Oh and even some more) but with unique arts and skills in the style of cards 😊
Also nice picture of treasure chest!
"Lady Nostalgia"
Wow! What a name! To make a model with that name I had to cancel my plans ! https://huggingface.co/Yntec/LadyNostalgia - thanks for the inspiration!
I also love making these, it has became some kind of obsession or addiction of mine, I had tried to stop or to take a break, but when I find myself playing chess or another game, I keep thinking "wait a minute, after the game is over it'll all be gone, I could be merging some image model instead!" and back I go.
Making TCG cards sounds so cool, I'll keep that in mind when releasing models that could be useful for that, along with others that merge the Gacha model into them!
Hey @Yntec , I'm really happy to see that you created a model with Lady Nostalgia name, lol 😂, That's an honor from yourself : your self-made models are incredibly amazing and creative, I love many of them!
And seeing a model that's named by the term I just used in my previous answer is an unbelievable surprise! You made my day and I will test it very soon with a lot of my self-written prompts notes!
SOO THANKS ❤️
🎄🎁 It's christmas before time 🤭!
Hahaha yes, I understand you very well, playing with the infinite possibilities of AI Generated Images, building new possibilities can be really a lifestyle sometimes.
My favourite chess piece is the Queen due to her moves freedom!
If you make a TCG game models, may I suggest you to bake together at least :
Incredible World series models + Gacha + Lady Nostalgia?
What will become from that cooking recipes?
Hehe!
:D I hope you like LadyNostalgia, I'm a fan of her style myself, here's a comparison with Gacha:
an illustration of a baby hedgehog with headphones holding a bow ribbon umbrella in the rain
Euler - Steps 21 - Seed 17971 - 512x512
Gacha on the left, LadyNostalgia on the right.
But I wouldn't know how to build a prompt for a TCG graphic to compare.
What will become from that cooking recipes?
Unfortunately Gacha includes a model with a license that doesn't allow me to share merges I make with the model, so what I did was to not share its recipe, that way they can't go after me, lol! I learned my lesson after having to take down the potatoMash model: https://huggingface.co/Yntec/PotaytoPotahto/discussions/2
But what if to compensate I don't make you one, not two, but three models? Incredible World + Gacha, Gacha + Lady Nostalgia and Lady Nostalgia + Incredible World?! It sounds like a fun project! Perhaps one of the new models manages to be better than the existing ones?! Hopefully, it's an offer you can't refuse, haha!
So here's the first one! Incredible World + Gacha! https://huggingface.co/Yntec/IncredibleOdds - I made it yesterday, but forgot to announce it here, lol!
Here's the second one! Gacha + Lady Nostalgia! https://huggingface.co/Yntec/GrandPrix - I'm liking how these are turning out, one more left to go!
Hello @Yntec , I need to test a little bit more LadyNostalgia but what I can say for the moment, is that it's a n interesting model that is not ignoring prompt terms when they are many comma separated !
I like it yes. But I discovered that you labeled it as "PhotoRealism" but it's not very set to get photorealistic inferences, color is too much vibrant (which is certainly normal, since that model is using NostalgicLife that is a model more Anime oriented than PhotoRealism oriented).
I don't say that it's a thing that it must be changed in THAT model as it is a very fun model to play with! 😁 But perhaps releasing a new LadyFabulysm model (per example) that will have photorealism priorities?
In other words I like LadyNostalgia, I see a lot of potential with it. But another model oriented Photorealism whith conceptual main keys that was putbin LadyNostalgia would be amazingly nice too, no?
I see in LadyNostalgia also the probability to get something crazy for some inferences, but well, hmmm... I'm going to test your model LadyNostalgia even more and giving more feedbacks soon !
I didn't tested the other two for the moment!
But I will do aswell, thanks greatly a lot for your support to my ideas !
(and yes I'm sad aswell to discover that Gacha have such of a license).
I'm excited to see what you will produce with your cooking talents that times!
I was able to produce these awesome LadyNostalgia images, that's a model for arts, for sure, it would be amazing in a TCG arts-display implementation :
(love the creativity and highly randomness, that's what I can state for Stable Diffusion 1.5 will reach high levels when needing creative and inspiring arts images !
Looks @Yntec :
Thanks for the feedback! You're right! I have removed the photorealism tags from the model! I have trapped myself in some sort of bubble, limited by the same prompts I always test, and then, I just note they look good, and then copy paste the tags from the models I merged, which can lead to such mistakes! I'll keep in mind making a version of LadyNostalgia that enables what I promised, it'll be high priority after I finish these models I've planned!
Haha yes, 🙆🏻♀️ quick copy/paste without editing them while testing are something what I know very well myself 😂😇!
Thanks for your thanks 😊!
From my side I will continue to make feedback on your other models which you merged with LadyNostalgia.
Finally for that part of your answer :
"I'll keep in mind making a version of LadyNostalgia that enables what I promised, it'll be high priority after I finish these models I've planned!"
I'm now very excited to see any news on that interesting series of models!
Keep on up with that good work, It will be also highly profitable to any people who believing in State of Arts by AI !
Again thanks and see you on my next feedbacks!
I really appreciate it because you're the first one critiquing my models! I had no idea Gacha was that good, I recall I wasn't able to make girls with good eyes with it so that's why none were present in the samples, heh, I could have been making merges with it much earlier!
So here's the third one: Lady Nostalgia + Incredible World 2! https://huggingface.co/Yntec/NostalgicWorld I really like the fine details on this one, though i can't say any new merge defeated the originals.
I dream with a world where everyone that can, grants the wishes of those that can't, recently many wishes have been granted to me so of course I'll continue making merges until you are satisfied! Grant the wishes that you can of the people surrounding you and continue the chain to make the world a better place!
Hello @Yntec !
Thanks for the discussion!
Gacha is a fantastic model, in terms of creativity for sure it rocks on the paper!
I agree that eyes modeling with Gacha seems to be having a falsely behaviors.
But for creating cards illustration it is a model that can be used for its unique creative skills!
Speaking of LadyNostalgia (vanilla version), it's just a awesome model which extends creativity to a higher level than Gacha is capable of.
LadyNostalgia is likely a Gacha 3.0 hehe!
I mean, LadyNostalgia is a fabulous replacement to Gacha, so incredibly creative (and I only using HuggingFace Serverless API to tests models, it's a perfect ways to make benchmarking more precise and on same options levels from one model to another one).
Plus: the license is more flexible!
I'm really astonished that you never had feedbacks on any of your models!
On HuggingFace you are a great model crafter!
(Norod78 too, with models that are totally different from usual. But even that they are interesting, as they use Stable Diffusion XL technology those are pretty uncreative.. Sadly).
I like a lot discovering your models, and as you publishing regularly models, I have always something new to play with!
If you want I can do some tests on your other models which are not merging LadyNostalgia in them.
Also I'm honored to know that you are making merges until we're (other people) satisfied.
But, even if it's not for me, trust me, here you currently have the best public Stable Diffusion 1.5 recipes!
You opened so much choices, in many themes. 😎
Aw, shucks! You say so many nice things, I don't know how to deal with them, haha!
Sometimes all I do is putting two models on the SuperMerger, ticking use MBW, save model in options, put a name, paste my magic number ( 1,0,0,0,0,0,0,0,0,0,1,1,1,0,1,1,1,1,1,1,0,0,0,1,1,1 ), see that it does what I wanted, and release! All the work was previously done by the original authors of the models, that's why I don't feel as an author and don't put myself up there, one day I was going to make a collection of my merges, but then, what was I going to call it, "Yntec's models"? But then people would think I made them!
I think the worst offender was Yntec/KomowataHaruka, all I did was pick up an existing model and merge a Lora into it, how could I name myself the author of it?? 😅
I have also gotten feedback from people like digiplay and Crowyote and appreciate it as well, I'm not keeping a list because I'm sure I'd forget someone!
It's great to hear LadyNostalgia lives up to her name! I just uploaded Yntec/GameIcons3D if you love Gacha, perhaps you'll like this one too as it's from where the compositions come from! I guess it's better than Gacha at doing objects, but worse at characters, though, the gacha mechanic is all about rolling and rolling until you get what you want, GameIcons3D would make a better pic if you rolled for 10 ones on both. Or, who knows? It's pure speculation since I've only made 4 pictures with this last one! 😰
Sorry for the late response, I had some stuff that needed to be done haha!
Hmm.. Well @Yntec , an author and a crafter is different terms.
In my viewpoint an author is initially the master of something that was created.
Where in my viewpoint a crafter is an artisan that build something new and refreshing based on something created by others!
A crafter is a passionate person where his/her goal is prior to improving something and building a different future together with author of something!
So, even if all you doing is merging stuff and cooking them with settings and so on, then, you're aswell building something new and refreshing.
Of course a crafter can't appropriate the works of others, so putting your name other initial models, even merged, is just not fair!
And you understood that fact very well from what you said, so authors be able yo exists through your crafts!
That's a benefits for them! (unless their licenses permit it!).
You're a crafter! Continue crafting amazing things and refreshing old models to became popular ones again! 😁
I will continue testing LadyNostalgia, but also continuing aswell discovering new well-crafted models from you!
I like digiplay original model which is called PhotoReal 2.0,but looks! Your crafted digiplay PhotoReal improved model is much better than the original! Of course your craft will not be possible to exists without the initial works of digiplay! But well you craft arts by improving their models! It's all good for the community and even for the initial authors of those models which you always ctediting them on yoir published models page!
I will look to GameIcon3D of your crafted models! Will feedback on it as soon as possible! 😁 Thanks!
Don't worry, I'm usually the guy that takes a very long while to reply, recently I replied to a four years old thread, lol!
I just released https://huggingface.co/Yntec/XP3D it's a model that uses Gacha's core compositions with 3DCuteWave's eyes and the C4D lora, I have become obsessed with eyes and tend to sacrifice other things like creativity for them, but while this model has huge gaps and weaknesses, I like the things it does well (I never liked 3DCuteWave's empty backgrounds, XP3D at least tries.)
I support your idea of being someone that crafts models, it's tempting to one day go and make a dataset and finetune my own model, though, maybe I'll wait until someone makes a creative version of flux.
Something that would help would be to see what kind of pictures you are generating and their prompts, and an example picture of an output you'd expect to get, I tend to craft the models based on some idea or trying to achieve something, but it's usually adding backgrounds or fixing eyes, or allowing a model to draw concepts it has missing from the other model. It's always something I see some model can already do, and I add it to other models, but I haven't really attempted to make them replicate some existing picture.
Hi @Yntec !
Sorry for the late response, life required me.. 🙃
Four years-old thread 😂.. Hmmm Kay'.
Currently I was needed to get me off from testing models, due to life pressure 😂. But yes I will be able soon to get back into testing!
If you're obsessed with eyes, perhaps you can look to your La-dee-dah model, eyes seems really impressive for characters!
And do some merge with that model !?
You seems waiting for a Flux advanced model... Hmm.. But, today, it seems Stable Diffusion 3.5 medium has been released by Stability AI, and is already available here at HuggingFace, at least for testing what it lies below the new hood haha!
Hmm, well, then perhaps instead of waiting for better Flux, looking for current new Stable Diffusion 3.5 Medium model?
Well, hmm, of course, I would be glad sharing images + their prompt for some of your models, I will be perhaps more organized by posting those images and prompt in page of each community threads on your own model pages! What do you think about this idea?
I already tryied to get a "same view output" with some of your models, aswell LadyNostalgia, for being honest I think we needs to have a specific StableDiffusion_1.5 image-to-prompt tailored to handle that hard task. Due to the fact that we can't really count on DeepBoru or simple ChatBot, legends maker, to build the real prompt to recreate a copycat image from original picture!
My tries related to do copycats give me a serious expression of deceptive emotions! But it's not bad, it's just too much creative by any Stable Diffusion 1.5 model (due to the operations in version 1.5 of Stable Diffusion), really too much creative towards the original pictures and what it describes inside!
Hey FashionStash, welcome back! I guess my problem with merging so many models is that eventually I forget what my old models were doing, I can't even remember what the outputs of Lady Nostalgia looked like without looking them up, ha!
If you're obsessed with eyes, perhaps you can look to your La-dee-dah model
The problem with models with great eyes is like they're like one-trick ponies, you can't really use them to fix the eyes of other models because all those models will start looking alike! Frankly, I think my favorite eyes ever come from... (well, they come from LexicaApertureV2 but that's a private retired model best left unmentioned) come from... LiberteRedmond, so some of my favorite models include that in the merge.
I could have actually released more models with better quality than I have, but they'd look very similar to existing models, it makes me proud to try the same prompt and seed with many of my models and see them produce radically different results, like Lunatic, 365, DarksideRemix, AliveAndKicking, XP3D and ReVive, they all give completely different outputs, I go out of my way to got that, but I'll keep in mind to make more merges with La-dee-dah, it's just that... the models I have planned to merge with it already got good eyes, hehe...
looking for current new Stable Diffusion 3.5 Medium model?
Doesn't seem better than the Stable Diffusion 3.5 Large model... apparently it exists for people that can't run the Large version in their machines, but I can't run either! I can't run any model with my hardware! Since I rely on online generation it makes no difference to me.
At the end of the day, Stable Diffusion 3.5 Large disappointed me on the sameface department, I mostly generate illustrations of cartoon girls that aren't anime, and while 3.5L does an outstanding job with them, they all have the same face... They, like Flux, suffer from lack of artists and styles because they got rid of them on their datasets and retagged them with an AI that didn't recognize them, so they're there, but SD3.5 doesn't know how they're called.
I think SD1.5 with an improved text encoder and VAE would be our best shot, because the creativity we love comes from unet... At least SD3.5 Large is more creative and varied than Flux, but is behind on aesthetics and Flux has more varied faces, SD3.5 Medium's only advantage over Large is being able to do resolutions different from 1024x1024...
I will be perhaps more organized by posting those images and prompt in page of each community threads on your own model pages!
Sure! digiplay used to do that and then I was posting his images on the Model pages!
Due to the fact that we can't really count on DeepBoru or simple ChatBot, legends maker, to build the real prompt to recreate a copycat image from original picture!
No need to do that, you can just post the real picture, then I'd see what I'd try the merge to aim at, perhaps for that I'd get in there some model we haven't mentioned yet because it's already close to the goal image!