[Stable Diffusion] Prompt Sharing and Learning Thread

[me3]

Jimwalrus i have done some test, and for what i see, 30000 step tend to do chromatic aberration at mid cfg.
so i decide to try different train, and i finish to get my best at 3000 step.
I share setting, maybe if you want to try.

15-20 images good quality, close-up and full body (i use 768x768)

Create embedding:
Number of vectors per token : number of image/2.2 (rounded at high value)

Train setting:
Gradient Clipping :norm
Batch size : 1 or 2 (depends on vram)
Gradient accumulation steps : 1
Embedding Learning rate : 0.005:100, 3.09:500, 1.8:700, 2.06:900, 3.269:1000, 1.05:1500, 0.06:2200, 0.9
Max steps : 3000

With these setting i get good result in about 30 mins, and usable TI from low to high cfg.

I made different ti for testing with same setting on different model, for what i see do a train with f222 model get better result for realistic.

Those learning rates seem insane. General advice seems to be to have a pretty low rate to (presumably) hit the vector weights you want and to maintain flexibility. LR like those might work very well for things you want want to keep very "fixed", ie like a style you'd not want to make changes to or a very distinct subject, like just the face of a person with fix expression/features.

If you'd use some of those LR just for a single Epoch you'd have your subject learned, so i'd imagen using it for 3000 steps on 15-20 images you'd beaten that data in really hard. So you'd probably have a good likeness to your dataset but would need very high weight modifiers to your prompts to change this, if you could at all.

(sorry about dragging up a 1,5month old post but reading through all pages takes a while, even just skimming through some of it)

 

[HardcoreCuddler]

Those learning rates seem insane. General advice seems to be to have a pretty low rate to (presumably) hit the vector weights you want and to maintain flexibility. LR like those might work very well for things you want want to keep very "fixed", ie like a style you'd not want to make changes to or a very distinct subject, like just the face of a person with fix expression/features.

If you'd use some of those LR just for a single Epoch you'd have your subject learned, so i'd imagen using it for 3000 steps on 15-20 images you'd beaten that data in really hard. So you'd probably have a good likeness to your dataset but would need very high weight modifiers to your prompts to change this, if you could at all.

(sorry about dragging up a 1,5month old post but reading through all pages takes a while, even just skimming through some of it)

LR all depends on errors afaik.
If the difference of errors (difference between what you get and what you want) between epochs isn't high (you're not making progress) you need a bigger LR.
If the difference of errors between epochs varies greatly and possibly randomly (you're getting all over the place results) you need lesser LR.
Keep in mind "bigger" and "lesser" can mean anything from +-0.001 to +-0.1 LR. You really don't know until you've tried everything, though usually these values are adjusted vvveeeeryyy slowly.
A very small LR isn't always good though, because it can get you stuck in what my prof called "pits" in the error field (I don't remember the dimension definitions of the field, but the Y axis were the errors). Basicaly, a higher LR would allow you to skip or get out of those pits. Still, the issue with high LR is that those pits may be all that's available in a situation, so you may be missing the global minimums (deepest pits, which may be very small in 'diameter') because high LR doesn't have a very good resolution.
That's literaly all I remember from my AI class lol.
TL;DR: Try a jumble of LR's until you run out of things to try, unless you know some super advanced math that not even my prof understands properly.

 

[me3]

LR all depends on errors afaik.
If the difference of errors (difference between what you get and what you want) between epochs isn't high (you're not making progress) you need a bigger LR.
If the difference of errors between epochs varies greatly and possibly randomly (you're getting all over the place results) you need lesser LR.
Keep in mind "bigger" and "lesser" can mean anything from +-0.001 to +-0.1 LR. You really don't know until you've tried everything, though usually these values are adjusted vvveeeeryyy slowly.
A very small LR isn't always good though, because it can get you stuck in what my prof called "pits" in the error field (I don't remember the dimension definitions of the field, but the Y axis were the errors). Basicaly, a higher LR would allow you to skip or get out of those pits. Still, the issue with high LR is that those pits may be all that's available in a situation, so you may be missing the global minimums (deepest pits, which may be very small in 'diameter') because high LR doesn't have a very good resolution.
That's literaly all I remember from my AI class lol.
TL;DR: Try a jumble of LR's until you run out of things to try, unless you know some super advanced math that not even my prof understands properly.

Your initiating text should come into play as well, with the basic idea that it puts you "close" to where you should be so you shouldn't need to do that much jumping around, More complex the subject the more complex a search though.
Given how many TI, Loras etc people are posting on sites that are having wide effects on things they shouldn't and are difficult to work with/combine, I'm worried that ppl see the quick results without considering side effects. So instead of "fine tuning" and carefully adding to the composition, it's just being brute forced in. Which might make it difficult to have things working together.

 

[me3]

Since i mentioned single epoch training:


Left is from the training set, right is the first image generated using a 1 epoch TI using a 0.9 (i think) LR. (prompt was basically just "<name> bikini location detailed face and eyes", not sure the detailed face/eyes was needed, i just didn't remove them from the prompt that was already there, being lazy and all that)

 

[Mr-Fox]

Her special treatment for those really bad itches..
Yes she is obviously Danish and veeery patriotic..

 

[me3]

Since i came across the kendra lora in here when trying to catch up with the thread. I think mine might be slightly broken, can't quite put my finger on it, but something seems a bit off....hmm


A couple of misses that potentially can be saved with some in/out painting:

Spoiler

You don't have permission to view the spoiler content. Log in or register now.

(sorry for any crushed dreams and nightmares)
(edited to link the post with the lora)

 

[sharlotte]

This Kendra:

You must be registered to see the links

Yikes, I'd say your prompts might need rework. Though to be honest, she of the LORA does not look much like Kendra Lust, but more like Lisa Ann on the images available in civitai.

Anyway, back to the spinning wheel (Ferris in this case):

 

[Mr-Fox]

This Kendra:

You must be registered to see the links

Yikes, I'd say your prompts might need rework. Though to be honest, she of the LORA does not look much like Kendra Lust, but more like Lisa Ann on the images available in civitai.

Anyway, back to the spinning wheel (Ferris in this case): View attachment 2640324

I think he might talk about my Lora of Kendra by SMZ-69. Good job on the undead nurses btw ME3. I'd like to spin those wheels Charlotte..

 

[me3]

This Kendra:

You must be registered to see the links

Yikes, I'd say your prompts might need rework. Though to be honest, she of the LORA does not look much like Kendra Lust, but more like Lisa Ann on the images available in civitai.

Anyway, back to the spinning wheel (Ferris in this case): View attachment 2640324

Mr-Fox is correct, to be fair i should have addressed my post and linked the post with lora.
Probably because of the thousands of images i've been reviewing (and discarding) lately when testing training (oh the cursing and swearing...) lately, but her left hand...seems she doesn't come up short in that area either , the ever returning problem.

I think he might talk about my Lora of Kendra by SMZ-69. Good job on the undead nurses btw ME3. I'd like to spin those wheels Charlotte..

Thanks, i'm just trying different things to see what works and maybe inspire ideas for myself and maybe even others. Even in just what you stick in a prompt

 

[me3]

When you're working with 2gb vram and forget to fix/set your prompts and walk away because generating is predicted to take a long time...
Happy accidents i guess (just don't start counting fingers, nipples etc, get distracted by other things...)

 

[Jimwalrus]

When you're working with 2gb vram and forget to fix/set your prompts and walk away because generating is predicted to take a long time...
Happy accidents i guess (just don't start counting fingers, nipples etc, get distracted by other things...)
View attachment 2641298

That's very impressive for being done in 2GB of vRAM. I couldn't manage anything like that with my old GTX960!

 

[me3]

That's very impressive for being done in 2GB of vRAM. I couldn't manage anything like that with my old GTX960!

tiled diffusion does most/all of the work, but you can't exactly be in a rush

 

[me3]

People are probably already aware, but i can't remember seeing it mentioned in the thread already.
By taking advantage of when elements/subjects are "removed" from a prompt or "transitioning from one to another you can create some nice effects. Just some examples and i tried to keep the prompts fair short and clean. What result you get will change a lot depending on seed so not sure i'd call it predictable, and the actual second element/subject might mean a lot less than you initially think, IE the simple green grass in the 3rd image.



(Edited to add some more examples)

 

[qazeqaze]

Have been messing about with:
Checkpoint:

You must be registered to see the links

VAE:

You must be registered to see the links

Upscaler:

You must be registered to see the links

Extensions:

You must be registered to see the links

+

You must be registered to see the links

+

You must be registered to see the links

Positive prompt:
Beautiful male, __portrait-type__, __artist-anime____hair-color__, __eyecolor__, __clothing-male__, __headwear-male__, (masterpiece, best quality, high quality, highres, ultra-detailed), __forest-type__, __flower__

Beautiful female, __portrait-type__, __artist-anime____hair-color__, __eyecolor__, __clothing-female__, __headwear-female__, (masterpiece, best quality, high quality, highres, ultra-detailed), __forest-type__, __flower__

Negative prompt:
bad_prompt_version2, bad-artist-anime, bad-hands-5, bad-image-v2,

View attachment 2635999


View attachment 2636000

View attachment 2636004
View attachment 2636033

That space/water image is amazing. Really, really like it. Nice work.

 

[Mr-Fox]

Revisiting an old prompt. This time I'm using my own

You must be registered to see the links

.

 

[me3]

Revisiting an old prompt. This time I'm using my own

You must be registered to see the links

.

View attachment 2646269

judging by your prompts you seem to have been trying to fight the pretty hard to win battle of models generally just thinking of "bikini" as one thing and not having any real concept of "variants"

 

[Mr-Fox]

Revisiting an old prompt. This time I'm using my own

You must be registered to see the links

.

View attachment 2646269

judging by your prompts you seem to have been trying to fight the pretty hard to win battle of models generally just thinking of "bikini" as one thing and not having any real concept of "variants"

That's very common with SD.
You kinda have to use a hammer and hit it over the head repeatedly to get the result you want..
SD doesn't know anything it hasn't been trained for. How would you even name or categorize different styles of bikinis in the first place. With SD your best bet is to try describe it instead of naming it. What ever "it" is.. With this prompt I was trying to get a small bikini, both bottom and top. The problem is that the breast size is not being differentiated from the size of the top.
With a huge rack you get huge bra or bikini top too. So I have tried to find a way to describe it. I used undersized bikini, it didn't work as intended or hopped for. I guess I'll have to include it in the dataset next time I make a Lora.

 

[me3]

yeah i've been clearing out thousands of images from when i was testing what kind of "common" naming/sizing/styling that would work.
Considering how much clothing/fashion images there are online it was a bit surprising how many basic concept hasn't been picked up. Just a basic/simple thing as using bra sizes would have been very effective and, you'd think, easy to do considering all the existing images from models, clothing stores, fashion magazines etc. Hairstyles is a pretty large hit and miss too, even basic hair lengths.

They hould have spent less time pissing of artists and more time looking at boobs and head shots

 

[Mr-Fox]

yeah i've been clearing out thousands of images from when i was testing what kind of "common" naming/sizing/styling that would work.
Considering how much clothing/fashion images there are online it was a bit surprising how many basic concept hasn't been picked up. Just a basic/simple thing as using bra sizes would have been very effective and, you'd think, easy to do considering all the existing images from models, clothing stores, fashion magazines etc. Hairstyles is a pretty large hit and miss too, even basic hair lengths.

They hould have spent less time pissing of artists and more time looking at boobs and head shots

Yeah I agree. We should all just look more at boobs..

 

[me3]

Look at that, she has a freckle/mole on her ass...