In my opinion he did a stunning work. There are several things AI trainings (even the most perfect ones) can't get landing and sometimes it can be symbols, other mix of several materials (without coliding nonsensing with others and so on), scars or marks, letters, numbers, etc.

Don't get me wrong I love your idea of entire ecosystem mixture in Zonai but as I see the things with today trainings, this probably is the 90% of what can you get (since always there is a bit of room to improve).

Being said that I have absolutely no doubt in the future trainins will be almost auto, easier and of course way way way more precise and coherent.

Good job Kony!

@LDWorksDavid I bet he could get it even better, but he certainly has no obligation to do so! I figured I'd offer some thoughts/tips in case he does decide to try any further, hehe. I think you're underestimating what's possible right now, it's just doing much better might be a lot of work. I won't blame him if he doesn't have the time or interest. :)

@ouros all good and I agree on everything dude except in understimating AI part. I'm not understimating it, I'm being extremely realistic. I have been training since the DD era and SD with textual inversions in VOC when you needed more than 16 GB VRAM for single embeddings/TI's like... 9-10 months ago before Dreambooth even existed so seeing all the available Zonai dataset in Google, knowing what Kohya_SS can do and dataset quality I really think you're overstimating a bit what Kony can do with this. Already said it days ago, finetunners are limited by AI it self and sometimes you need to "let go" in order AI to "give".

And of course all the tips, etc. awesome and super welcome for sure but knowing AI limitations also is fundamental. Cheers!

@LDWorksDavid Woah, sounds like you've been at it a while. I never even got Disco Diffusion working myself, haha, nor have I done any training for DD or SD.

Konyconi's tutorial suggests he uses almost exclusively synthetic data, though. My guess is that the choice of data could greatly change the result, and that mixing in more data with relevant aesthetics - and possibly crafting custom data to fill in anything it's missing - could improve the result. That's entirely speculation, though, and the bespoke custom data is an especially heavy lift, way more than I would expect Konyconi to do except perhaps out of the challenge of it. (Also, training a model to generate synthetic data to then train another model could be interesting....)

I guess what I'm saying is if you're already heavily using synthetic data, the sky's the limit? (Well, really the AI is the limit of course.)

@ouros yeah! Totally. With custom data is another whole story. If you can draw, paint or model in 3D you can vastly improve your results cause the curated dataset will be better. Usually we don't need to do it since AI gens outs, artworks in the net, etc. Are enough for most things (and it's more time consuming). The common limits there I can think is letters and hand/fingers. You're absolutely right here. That's not speculations :)

@LDWorksDavid Guess it all depends on what Konyconi is willing to do. :) But I would not be bothered in the slightest if he chooses not to pursue it further. If he did pursue it further, I bet he could do even better!

OK. I'll see what I can improve here.

@konyconi Let's go dude! I want to see what you really can do aside this... will be awesome to put the training images to actually show people sometimes which images we have to use to train some concepts. Cheers!!

@konyconi Yaaay! I look forward to seeing what you can do! :D No rush, though. Seems like you've already got a million things on your plate. (Which I greatly respect, but don't burn yourself out.)

@LDWorksDavid If you haven't already, go check out Konyconi's tutorial. You can download his training images for that style and take a look at them yourself!

To not only come up with effective and clever techniques, but to then share them freely with the community... Konyconi is truly a hero and a legend. Seems he thinks the simplicity of his process makes it less impressive, when I would argue simplicity is itself a virtue. A simple and effective process is -better- than a complicated and difficult one.

Oh Konyconi... in the result of your labors, my hard-drive weeps but I rejoice.

@ouros He knows the tutorial very well -- he managed to guess most of its content before it was published :D

@konyconi Well, 10 months training gives me some field experience (even I'm "new" to a lot of people here). Still I really think the Bing image addition gives something very specific in terms of "noise". I suppose you have already figured out but if you train from one single source it's not the same (theorically worse) than training from a mix of multiple sources. Bing was a thing I didn't expected (and same with cosine with restarts, I expected contant or cosine). The other thing I didn't excepted was the color augs, I was expecting a Noise Offset of 0.05-0.2 or even a multires. Indeed was very cool seeing your settings.

@ouros Yeah, I already checked it the first day came out. All logic but the Bing thing was like, "really?", I was thinking Google or MidJourney mixing but not Bing, haha. That was unexpected, Since I saw him doing the textual "punk" inversions already knew what his most impressive tool was, his brain. The way he pull styles almost from a minimum concepts it's insane. Already told him, that he was putting himself way bellow where he should be. So yeah, very cool seeing him at top ^^.

@LDWorksDavid @ouros  Guys, check the pictures in the description -- that's my attempt to add animal motifs.

@konyconi Oh wow, that's fantastic! Looks waaay more Zonai already! I like the lower-right one especially.

As far as further improvements, a lot of the Zonai animal stuff is pretty abstract. Like Aztec animal sculptures and carvings, they can be kind of boxy and stylized. I didn't even realize some Zonai techs were stylized animals until I went looking for it. So just a little bit less animal might be good. You might also consider throwing a bit more "Aztec" into the mix.

(The temple to Quetzalcoatl at Teotihuacán may have been used as inspiration for the Zonai, so you might find that interesting to look at. There are two repeating animal head sculptures on the sides, one of which only vaguely resembles an animal at all.)

Also - and this is kind of nitpicky - but I feel like the blue-green color is a little too blue. It's meant to be a kind of bluish jade I think, though jade varies in color and what you have already is not impossible for jade. No worries if you leave that color as is, just thought it was worth mentioning. :)

You're a wizard, Kony. An AI style wizard. Thank you for your continued effort!

@konyconi Looks great. I don't have a deep Zonai knowledge but I judge from the dataset available. Good job. If you want to follow the less animalistic aesthetics is either or reduce the symbols/motifs/representations in terms of deleting images, zoom out then could work (of take out flip aug) too or tweak the "dangerous" Dim/Alpha. I suppose you're at 128, 128. With the augs enabled I'm not sure till which value you need to drop to decrease the effect but I'm certain of one thing. Don't let the alpha goes too high. That's all I can think aside of noising in settings. About the colors "fix" probably caused by combo BF16+augs (since happened in your other LORA if you remember). But from what I see, still it looks like a very good take.

@LDWorksDavid Seems that we are on the same wave :) I've just switched off the color aug, and cleaned the dataset.  

OK, this version (v.8.0) is final. I don't think I can improve it more. Tried to make animal motfs more abstract, didn't manage to fix te color. I'll share the dataset later (there, the color is more green-is and less blue-ish). @LDWorksDavid : If you want, you can try something with it.

@konyconi To be fair I have a lot on my plate in terms of experiments/testings and embeddings making but will take a look. sure. Being said that, seeing your outs I think I really can't do it much much better U^^. Probably they will look similar. Still is a thing I could try in the future so thanks for the datasets mate. Take care!

@konyconi Wooaah. To be honest, I'm not sure I expected this much improvement when I initially commented. Or maybe it just looks so much cooler to see than I anticipated. Either way, I'm absolutely thrilled. :D

Thank you so much! I LOVE how this looks. You're truly a wizard, Kony. I don't know how many of your LoRAs I've downloaded at this point, but you just keep making amazing things.

@LDWorksDavid And you thought his first attempt was nearly the best we could get.... You underestimate Kony! (...and the maaagic of synthetic data.) But to be fair, it's hard to grasp the full potential of the AI we have right now. Who knows what's possible with the right data. I've never seen something in technology that felt quite as much like arcane rituals (or consorting with demons...) as working with AI. (Though coding deserves a mention here.)

@ouros Yeah, and I wasn't wrong and didn't underestimate Kony at all, lol. I consider him a good finetunner and a extreme masterbrain. I think you didn't understand my point mate U^^. That was the best he can get BUT with a non edited/curated dataset.
Already explained why the second method is a HUGE improvement always. Takes more time but leads to precise results but what I meant was always from an original data you can get from images in google, artworks and so on -or gens- (which is what we usually do). If you start cleaning parts, drawing here or there or tweak it via "zoom in, out, crop" then that's not the same dataset.

Anyways, all is clear and the LORA is looking good ^^. I got your point and I also admit that sometimes AI it's out of this world in terms of outcomes.

About the limits of AI, that's subjective but I think its 70-80% input source+20-30% training/curate-edit.

@konyconi Now we wait for SDXL to drop public and see if it's tuneable (and pray god for less than 24 GB Vram) :).

@LDWorksDavid  Regarding [ "dangerous" Dim/Alpha ] : I've looked at the math behind LoRAs and I believe that setting Dim == Alpha is quite justified. Or do I miss something?

@konyconi Maths shouldn't be always the way to go. To put it a quick example the loss rate theorically the less, the better. I reached 0.04 and it gets terrible results. About DIM and Alpha is more a global look. You need to look them but with Unet+Epoch (and LR restart/power) on the formula cause its not the same 128/128 on a singular condition than with other settings (and with augs change again).

So maths works in default scenarios because when you start altering variables like Unet, epoch, restarts and so on, changes the thing. That's why is not 100% a safe thing. For example, some images due to the "imagery" will work better on 32, 64, etc. than 128. Some LORA's works at 8-8 and others break in terms of consistency. Already said it but more than settings is WAY better to focus on getting as best as possible quality input images.

Of course is good to know that Alpha acts like a "brake" or a stop in terms of the Dimensions. A lot of people lose insane amount of time with settings and 2000 tests when the reality is that if you change a single variable, you can't do the same prediction (and I'm telling this as one of those).

In a nutshell, to have an overall global value 128/128 should be the way to go. If you want the best you can with the inputs then that's not always the case. More or less this.