SD: AI LoRA Training (Parameter Tips & Resources)

[@TheRoyalSkies]

For anyone who wants the exact dataset and json file used in the tutorial, you can get them from the doll on my ArtStation here completely free! www.artstation.com/marketplace/p/VOAyv/stable-diffusion-3d-posable-mannequin-doll? Also, don't fret to much if you don't really understand a lot of these settings, from what I've seen, 80% of the most important stuff is actually your dataset. The parameter part is 20% - I personally don't know most of them either, but they work none the less - Also, if you're having trouble, definitely drop by our discord under "AI-Questions", and we'll do our best to help out when we can! discord.com/invite/TdkyE9yZZ4

[@eddyxc1697]

the GOAT himself as posted. fr though, as much as some people hate AI, it's so helpful for people like me lmao. I'm a broke programmer who wants to make his own game and art/assets, but I dont have the time to learn how to draw.

[@IkoIkata]

Art for free at the expense of artists. That's okay, though, I prefer paying customers anyway.

[@darrennew8211]

I haven't followed along, but I really really should, because this is so cool.

[@sinayagubi8805]

Hey!! since when?? cool to see you in the AI space as well.

[@anonme2348]

>Cool bro, why don't you hire an artist? He is the artist >You are taking work away from artists He is the artist. >Just pushing a button doesn't make you an artist. No, the two games, tutorials, and the ton of art he's put out on art station makes him an artist. Man if you do a control net tutorial of turning your renders into 2D images, I expect the comment section to be filled with "just make it yourself bro". Okay, yeah, I should have skipped the rant went with "algo"

[@omnomnomynous]

I think there's more to AI backlash than that tho? like you say that no job is being taken away, but, consider that people without artistic experience see tut videos like this and think "oooh free easy art!" and don't consider the repercussions of mass-scraping data to train these AI. Guarantee they're going to be sorely disappointed when they realize you can't copyright AI-derived work LOL

[@lefourbe5596]

​@@omnomnomynousand yet i was able to do it anyway. What i generate is sometimes needed as part of a bigger products that is sold. Sometimes an on demand Lora for an ARTIST to use, i don't sell the pic but the model to help him complete manual work. I don't need to copyright my "lazy" work to make money. I just need to fufill a demand.

[@IkoIkata]

​@@lefourbe5596Art for free at the expense of thee.

[@caryonplays9024]

Always right to the point, thanks.

[@vi6ddarkking]

Considering its improved prompt adherence, I really am curious to see just how SD3 based Loras compare to the existing models.

[@froztbytes]

If you think about it none of us know how big SD3 is going to be. For all we know we might not even be able to train it on commercial graphic cards.

[@vi6ddarkking]

@@froztbytes You obviously haven't been around Image generation for that long. Thats the ENTIRE point of SD. SD3 will be released in several different sized models. So that everyone from the insanity of the RTX 5090 to the humble 2060 will be able to use it.

[@lefourbe5596]

assuming average PC have 8GB of nvidia card of course. you can lower this requierement to 4Gb with theses extra tips but it will hurt the quality sometimes : choosing not to train the text encoder (--network_train_unet_only in extra parameter of advanced tab). also you can use the experimental full bf16/fp16/fp8 for drastic VRAM improvement (lower accuracy) also Adafactor optimizer (instead of AdamW used here) will save on VRAM. lowering the network rank to 32 or even 16 or 8 will also save on VRAM; SDXL training might need 12Gb of VRAM with theses params. but i think it can be lowered to 8Gb --- for AMD owner it's difficult to run Kohya without linux. However on windows ZLUDA is making progress at converting the native CUDA script. it's janky but it would get beter.

[@meggriffin4395]

ill try for sdxl

[@llbsidezll]

People used to argue that using photoshop wasn't real art. Whenever a new technology comes along, there will always be people that are so attached to the way they learned to do something that anything that upsets the staus quo is seen as cheating. This is just a new tool for artists to add to the kit, not a replacement. The best thing you can do is learn and adapt as this technology isn't going away and will only become more integrated into professional workflows. It's powerful and way more efficient in helping go from concept to final product. Anyone learning to adopt new tech into their workflow will always have an advantage over someone who refuses to learn it.

[@IkoIkata]

Your argument doesn't hold up. Crappy Photoshops are still crappy and obvious. Photo bashing a subject instead of just photographing and drawing it is noticeable and garrish, just like the samey soulless AI crap that creatively defunct people put out. People will always be better. No neural network image output can stand toe to toe with a human artist. Anyone can see it from a mile away. Just learn to draw bro.

[@llbsidezll]

​​@@IkoIkatabut a crappy photoshop job where someone didn't take the time to learn to use the tool properly vs someone who has years of experience in photoshop are 2 different things. This is basically saying that no post work should be done at all, if you were really good you should be able to capture the image perfectly in camera? What about Adobe Illustrator? Is someone who creates 2d vector art not an artist because he uses the path tool to draw lines instead of drawing by hand? These programs were originally frowned upon as well but over time artists learned to adapt them into they're set of tools. An AI bro who just types a prompt into a field, hits generate and calls it a day is not the same as what's being discussed here. This is getting into how to properly use a tool and fine tune it to work for your specific needs. Just like any software. I'm not pro people selling AI art outright, but if these tools can be used to expedite the iterative process to help guide you to create something yourself, I don't see anything wrong with that.

[@llbsidezll]

I'm not an expert on this stuff. I'm here for the unreal, blender, and substance painter tutorials but I'm not angry that this topic is being covered on this channel. I'm not even using AI in my workflow atm but it's nice to know one day, if/ when the time comes that I need to learn this stuff there are clear and concise tutorials like this to make learning it easier. Plus I think it's genuinely interesting. As far as the AI stealing other artists work, I might be wrong but from my understanding this is an ai model that is only trained in your own data, not the works of others. So you can't add the term "in the style of H. R. Giger" and get something based off of his work.

[@lefourbe5596]

​@@IkoIkatawhat is beter is an ai assited traditionnal artist, i saw somes in projects under NDA and they are too good for me. they don't do txt2img they go back and forth with PS or other to dim the rough edges mades manually and vice versa with img2img. they don't sacrifice their creativity coz they have made their sketches before. they fill props that are not worth drawing, some of the extremes perspective are still made manually a well as stylized face expressions. most of them agree that it sucks because it can only do simple things but it suit them since they don't wanna do the simple things that hog time.

[@Dr.Fluffles]

Yep, what he's creating is a specialized sub-model which creates only what is fed to it when summoned by a code word or set of words, but with the final render using the knowledge of the base model it is connected to. Lora's can work with different models, to varying levels of success. There are also new models being worked on that would be entirely public domain and opt-in from the ground up, like Mitsua, which is being worked on for a specific Vtuber atm, but currently leads the pack.@@llbsidezll(To clarify, it won't pull from other people's influence on the model unless the base model would already be pulling from it, like if they prompted for them already on a base model which had trained on their work.)

[@juanjesusligero391]

Algo means something in Spanish!

[@kenhiguchi2144]

New ai learning video. Yay

[@SKOP320]

Another big Kohya ALGO comment

[@froztbytes]

And now we wait.

[@lucasouzah2a]

Is it possible to train the same data package or do I just have to modify the tags and redo the training? For example, if I wanted to improve training, how would I do that? Thank you for your videos and you do a good job, much success to you!

[@Mega-Mat]

This is something that I have been tempted to toy around with for a while now. Your video has made the whole process very easy to understand.

[@jawadoumar]

wait? For how long? Superfuckinf excited 😁

[@juanjesusligero391]

Will you upload the resulting model to Civitai? That would be nice for comparing with our resulting Loras too! :)

[@neozyykun]

can u make a tutorial on how to take lora's and pose them in t-pose for concept art? i tried it and it works but not very good

[@lefourbe5596]

Mmmmh i know how to do it but it range from kinda easy to fcking hard depending on design. I'm not sure on how to tackle an all around workflow for us to use. I try to reproduce my best 3D reference (99% usable for sculpting btw) with comfyUi using SDXL instead of 1.5 but with a worse openpose and the absent Tile controlnet it is tricky. I could even generate an img2img on top of the 3D reference sheet to make the topology.

[@sinayagubi8805]

Maybe also make a tutorial for Leco as well... I learned how to do Leco training properly from a Japanese video. there was no english video teaching it properly.

[@lefourbe5596]

Not4talent AI Channel. I know this guy well and he made a LECO slider for expressions and show how it was made. kzbin.info/www/bejne/fZK5poR7jNZ4pa8si=y3U8mOndA_Brbx6U

[@JadeEmperorTales]

What happens if my character is a creature like a minotaur?

[@allgreatnike1009]

Havent commented in a while. For the algo!

[@EthereallMoonlight]

algo

[@jawadoumar]

Dream of 3d artists who never learned traditional drawing

[@apoc519]

This content doesn't seem very popular and is instantly outdated. I'm not sure why you bother.

[@lefourbe5596]

I asked him sort of. Peoples Ask me all the times how i train my Lora and this video show exactly what i do with my preferences. Usually i explain them on discord for 15 to 40 mins. That is why i pushed skies to do this series instead of me making lenghty uncut video. Now i don't have to answer but all i need i to link this video. That is why

[@IkoIkata]

It's trendy and unregulated. He'll go back to legitimate game dev stuff when the AI imagery hype dies down.

[@lefourbe5596]

Of course he will that's expected. this is no blender guru one software channel.

[@IkoIkata]

To viewers: yes this is all very cool, but have you considered learning to draw? Or, god forbid, paying an actual artist?

[@dsweetz9236]

@IkoIkata Cool response, but have you considered learning how to create models in Unreal? Zbrush? Or god forbid, an actual artist use something to expedite their own workflow.

[@TheRoyalSkies]

Sure thing - Yes, I'm very comfortable creating both 2D and 3D works. This just helps me speed up my workflow; especially when I'm just drafting ideas for fun with friends or personal projects -

[@MC-qu9jw]

You are leaving this crappy comment on a video by someone that has made hundreds of tutorials on technical and artistic skills involved in game production, as well as free resources, they are imho the best tutorials on their subject matters on youtube. They have addressed concerns on ai, and have now moved on to cover what is an important subject in their field. TLDR; They are better than you as a human being, go be shitty to people somewhere else.

[@dsweetz9236]

@@TheRoyalSkies I'm on your side. I see that when I hit reply it didn't go to @IkoIkata. I do pretty much the same thing as you do. I use it to speed up my workflows for stock images for webpages that I set up.

[@TheRoyalSkies]

@@dsweetz9236 My reply was directed at IkoIkata as well - It looks like the @ didn't get automatically added for some reason - Regardless, yeah the way I see it, it's just saves so much time, allowing us to focus more on the aspects of the art we enjoy -

[@EmilyVanSchaik]

cringe.