Stable-Diffusion: Kohya Simple-Captioning (FAST!)

[@TheRoyalSkies]

EDIT: I have learned a bit more since uploading the video, and have actually found out that it's better to have a variety of different backgrounds, so if you have no green-screen then it's really not a big deal and probably OK - Next video, I'll show you how to do "manual" captioning, since I think it's good to be able to do both - Quick note, at 3:00 - I said "The simpler your caption, the more flexible the training can be" - But, what I should have said was "The simpler you keep it, the more *easy the training will be" - We'll talk more about this in a later video, but if you need any help, be sure to check out our discord under "AI-Questions" below! discord.com/invite/TdkyE9yZZ4

[@Rin8Kin]

>Subscribed for blender tutorials @ >learning stable diffusion 🤣

[@lefourbe5596]

same ! me love that :)

[@kenhiguchi2144]

Stable diffusion learning is fun.

[@Unknown-os8nb]

The only useful tutorials bc of my short attention spans are your videos.

[@lefourbe5596]

keywords : don't think too much about name, make a word that is short : Yra, Fxy, Syw, ahra, >>>> (the longer the word the more tokens (=space in prompt) it takes and therefore the more weight they will have compared to other words) the keyword should be trained next to the class of subject that you train like : a woman Yra, OR Yra, woman, etc.... /// the first word would get transformed first. THEN you input the background description as we want to separate it. idealy we would want to have a different background every time so the word green doesn't get affected but it would take too much time and results have shown to be sufficient on my end. usually, we train a close description of the whole character so prior knowledge get transfered to the new data. but it is a little more tricky to prompt after that.

[@kernsanders3973]

Recommend rather do SDXL, I would suggest ponyxl v6. Its pretty great, the coherence to prompts are light years above anime sd1.5 models, you'll need bigger resolution datasets. But thats honestly a good thing when comparing 512 to 1024

[@lefourbe5596]

i would backup this but it's a harsh experience for starters. i personnaly advised Skies to not go that way (mainly because he have a 3070Ti yikes) the hardware requierement is TOUGH and it takes 4 times longer, it's way easier to get your teeth and parameter done on SD1.5. to try things out make progress and understand what works and what's not before you buy a new PC. don't forget that we want ease of access and there is some Workflow that require SD1.5 controlnet, Tile resample ... i miss you... BTW i am using PonyV6(training) with AutismMix.confetti(generation) it is like you said THAT much more potent and you would see some examples in the series (if i take time to make some).

[@nocturne6320]

Tbh ponyxl is in my experience harder to prompt, individual characters seem to be too tied to their source artstyles and it doesn't seem to be able to handle anime artstyle that well, as it constantly sticks only to a specific style and no amount of prompting makes it switch to a different anime style I want. Compared to that, there are many SD 1.5 models that easily listen to style instructions

[@lefourbe5596]

@nocturne6320 l The finetune is so agressive that the base data got mostly erased somehow. there is many ways to counter it but it's not perfect. 1) use merged/mixed models. autismMIX is Pony based and Lora compatible and it's easier to prompt 2) one person have already made classic "styles" Lora for it. 3) "refiner" is a great option to use allowing to switch models during generation Ah... my answer for 1.5 usage (tutorial wise) got deleted, thx YT.

[@kernsanders3973]

@@nocturne6320 Havent had that issue personally. Styles I would tend to influence with Loras, score system and prompting together. Thus far havent struggled with characters in different styles. But when it's a character lora, then yes, there would usually be a style that would tend to come through if the weigh of the lora was high. About everything else I do with the base model. Even most concepts. Although have tried out a few concepts and didnt have much of problem to apply it to characters and different styles. Do suggest to try more of the pony mixes out there than just base pony. Think a lot of people give up to easily and go back to SD1.5, myself included at the start. It was only one day when I was reloading a previous prompt and trying to generate it again. Then all of a sudden a style/concept and character I was struggling to do previously with various loras generated perfectly. Thought it was a fluke and generate another and still generating it perfectly. It was then I realized I didn't have SD1.5 loaded but actually ponyxl. Not only that but realized none of the loras I used on SD1.5 was loaded. Meaning I accomplished it all with the base pony model alone. Not only that it was doing it perfectly. Thats when it finally really hit me how powerful and amazing that model was. Since then its my go to for generating. I would only switch to SD1.5 if there is a specific char/concept or style that I only have SD1.5 loras off.

[@nocturne6320]

@@lefourbe5596 I know the struggle with deleted comments... :/ But thanks for the advice!

[@terjeoseberg990]

I’d say that it goes through the images in that folder 20 times per epoch.

[@Lv7-L30N]

gracias

[@SKOP320]

Captioning Algo comment

[@enescelik1845]

i want to train a stable diffusion model like with 1800 pictures but its very slow how can i solve it ("20_modelfolder" epochs 10 ), it gives me 23-90k steps and its very slow

[@softsmolflower]

Not sure if I missed something, where do we get the training images?

[@verse369]

[@weaze5583]

cowboyra 🤣

[@Mente_Fugaz]

What hapenned to blender bro... AI is theft, we wanna make our own stuff

[@froztbytes]

Looks like someone didn't watch the Dataset Background Training video.

[@raxo7546]

algo

[@celestialx8]

Full of scammer these days using ai

[@FeralBlade]

A scam is a deceptive scheme or fraudulent activity designed to trick individuals or groups into giving away money, personal information, or valuables under false pretenses. So please explain, i will wait.