[Stable Diffusion] Prompt Sharing and Learning Thread

[Mr-Fox]

I've been on this

You must be registered to see the links

model kick, it even replaced Clarity for a quick minute:

View attachment 2686138

Holy ! something.. Lost for words and this doesn't happen often believe me.. This is definitely your best work yet.
Amazing , I gotta try this one out.

 

[Dagg0th]

Holy ! something.. Lost for words and this doesn't happen often believe me.. This is definitely your best work yet.
Amazing , I gotta try this one out.

You must be registered to see the links

I as well jumped into the A-Zovya photorealistic wagon.

For my first try, I manage to creep into this wench and took her a candid pic.

Gonna keep stalking her huahuahua..

 

[Dagg0th]

Does anyone old enough to remember the DOS game Dune from 1992?
Chani was one of my early crushes


Well this is the remake from 2023

Spoiler: spoiler

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

 

[Mr-Fox]

Does anyone old enough to remember the DOS game Dune from 1992?
Chani was one of my early crushes
View attachment 2687439

Well this is the remake from 2023

Spoiler: spoiler

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

Great work.

 

[Sepheyer]

So, with Clarity I stopped using "Plus Size Model" because the results were straight out of WALL-E. But with Zovya's, "PSM" yeilds sweet-sweet curves:

 

[Mr-Fox]

I was reading PMS at first..

 

[Dagg0th]

Hi guys, I was testing new merchandise last night, this one is a little feisty but maybe I went a little overboard; anyways, I will hear your offers...

Status: Used

Spoiler

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

Condition: untrained, unbranded, unbroken.

 

[FreakyHokage]

Where did you find them? Do you have the original png files? Those ones don't have the metadata you need to get an easy route to copying / cloning them
Interrogating the image though gives us this for the prompt that maybe you can use in a number of different Civitai models to see which fits best:

a sexy woman in a blue costume posing for a picture, hyper-detailed and realistic, fantasy dungeon, plus size woman, doll, hyper-realistic lighting, dark golden hair, featured on amiami, trending on artstartion, real-life big mom, blue skintight closed longcoat, ellora

--
Refining that with a bit of trial and error (we can do this better if we keep going), we can have this:

Positive: (masterpiece:1.2), (best quality), (intricate details:1.2), (photorealism), wallpaper 8k, a beautiful woman, wearing sexy revealing detailed dress, hyper-detailed, fantasy ornate castle window view, cleavage, wide hips, large breasts, sexy cloak, necklace ornament, (sheen to her skin), vibrant colors, (dynamic lighting:1.3), (vibrant cinematic lighting:1.3), (glowing lights and effects and lamps:1.2), lustrous golden hair, realistic skin, smiling, looking at viewer, facing front <lora:hipoly3DModelLora_v10:0.6>

Negative: poor quality, mutant hands, ugly, blurry, low quality, bad-artist, EasyNegative, facing front, boring, dull
--

That gives us something like the attachment - which we can see needs more lighting, and more colors... so you just keep adding to that in the prompt until you get something you need. If you have stable diffusion / automatic1111 - just throw the png into the png info tab and you'll get all the generation details.


View attachment 2531032 View attachment 2531040

If you want to get a decent prompt add the images to img2img and click Interrogate CLIP and it will give you details of the images. Also, click Interrogate DeepBooru and it gives you porn prompts.

 

[Mr-Fox]

If you want to get a decent prompt add the images to img2img and click Interrogate CLIP and it will give you details of the images. Also, click Interrogate DeepBooru and it gives you porn prompts.

It's a very good way to get started on the process of making a good prompt but it's very basic and needs a lot more. I haven't tried the DeepBooru option yet, thought it was only aimed at anime.. I will try it next time I begin from scratch.

 

[ririmudev]

Phew... made it through page 50 of 83. So far, seems like raster-image-oriented than vector-image-oriented, but well done (I think my brain resonates well with abstraction for whatever reason; I enjoy 8-bit music, pixel graphics, abstract algebra, etc). And shame on Seph for propagandizing us with thiccness

 

[FreakyHokage]

It's a very good way to get started on the process of making a good prompt but it's very basic and needs a lot more. I haven't tried the DeepBooru option yet, thought it was only aimed at anime.. I will try it next time I begin from scratch.

I did the deepbooru and it gave me this for the pink-haired girl. 1girl, blonde hair, breasts, brown eyes, cape, cleavage, curtains, dress, earrings, indoors, jacket on shoulders, jewelry, large breasts, lips, lipstick, long hair, looking at viewer, makeup, necklace, smile, solo, standing, window. Though the images reminds me of the Perfect World checkpoint.

 

[Mr-Fox]

DeepBooru is much more extensive than Interrogate Clip. I'm working on a prompt right now that started with DeepBooru. Thank you for the tip.

The office hoe aka secretary..

I'm having trouble with getting the fingernails but when it's done the prompt is great.

 

[Sharinel]

DeepBooru is much more extensive than Interrogate Clip. I'm working on a prompt right now that started with DeepBooru. Thank you for the tip.

The office hoe aka secretary..
View attachment 2691795
I'm having trouble with getting the fingernails but when it's done the prompt is great.

In my limited time playing about with both, it seems that interrogate clip seems to give out a sentence eg "A busty young woman in an office" while deepbooru seems to just have a lot of words seperated by commas "1girl, cleavage, office".
I like deepbooru for showing me some prompts that I didn't know about, there seems to be a lot of 2 word prompts that it shows as one with an underscore - deep_cleavage or the like

 

[me3]

Continuing my long running testing of training i've at least seem to have gotten my training times down which will speed things up. Tweaking just settings and optimizers the times have been from just 25min to 14hours on the same dataset... To give the process a more smooth learning it seems to have settled around 30-40 min now (batch 2 and ~2.8 s/it, which is considerably better than the 6.8s/it things were hitting at one point, if only i had a proper gfx card)

Will share some settings etc later once tested some theories to better cause confusion. Been training TIs as usually but considering the settings in kohya is basically the same for TI and lora a lot of things should be usable for both.

Anyway a small challenge. These are from 3 different trainings. All are trained on the same training images, same model, same batch/LR/optimizer/opt args etc. Generated on the same model as the training was done (no vae cause i forgot to turn it back one ), same seed/step/cfg/prompt etc.
So the question is, what do you think is making the rather large differences? Some things you can ignore, it's not due to lighting in the images, different epochs in training as the same "differences" was visible in the different epoch TIs and in multiple generations, even in the training sample images.

 

[Mr-Fox]

Continuing my long running testing of training i've at least seem to have gotten my training times down which will speed things up. Tweaking just settings and optimizers the times have been from just 25min to 14hours on the same dataset... To give the process a more smooth learning it seems to have settled around 30-40 min now (batch 2 and ~2.8 s/it, which is considerably better than the 6.8s/it things were hitting at one point, if only i had a proper gfx card)

Will share some settings etc later once tested some theories to better cause confusion. Been training TIs as usually but considering the settings in kohya is basically the same for TI and lora a lot of things should be usable for both.

Anyway a small challenge. These are from 3 different trainings. All are trained on the same training images, same model, same batch/LR/optimizer/opt args etc. Generated on the same model as the training was done (no vae cause i forgot to turn it back one ), same seed/step/cfg/prompt etc.
So the question is, what do you think is making the rather large differences? Some things you can ignore, it's not due to lighting in the images, different epochs in training as the same "differences" was visible in the different epoch TIs and in multiple generations, even in the training sample images. View attachment 2691939

I love this kind of experimentation. Looking forward to the next update. Good work.
Without know the settings and seeing the source images it's very difficult so speculate about what causes the difference. The only thing I can say is that language to image generation with "AI" is very unpredictable and inconsistent. Even if everything stays the same without any variables the result will still be slightly different. It's very much like rolling a pair of dices, you never know what you will get. If all the settings and images remains the same, then it's only the prompt for each image and how the model interpret said prompt that can be the cause for the difference. So if you wish to hone in on more predictable result then go over the prompt for each image and make it as clear as possible and the least subjective. I'm talking about the interrogate prompts describes each image and is used as source data not the sample prompt used during training to generate sample images. I hope I explained clear enough.

 

[Jimwalrus]

I love this kind of experimentation. Looking forward to the next update. Good work.
Without know the settings and seeing the source images it's very difficult so speculate about what causes the difference. The only thing I can say is that language to image generation with "AI" is very unpredictable and inconsistent. Even if everything stays the same without any variables the result will still be slightly different. It's very much like rolling a pair of dices, you never know what you will get. If all the settings and images remains the same, then it's only the prompt for each image and how the model interpret said prompt that can be the cause for the difference. So if you wish to hone in on more predictable result then go over the prompt for each image and make it as clear as possible and the least subjective. I'm talking about the interrogate prompts describes each image and is used as source data not the sample prompt used during training to generate sample images. I hope I explained clear enough.

I'm reading me3's post differently - I think it's a challenge for us, not a request for some kind of information or assistance.

Unfortunately though, as Mr-Fox says, SD is so unpredictable it could be anything (even though your post excludes most of the usual options), so I won't even try to guess!

 

[me3]

I'm reading me3's post differently - I think it's a challenge for us, not a request for some kind of information or assistance.

Unfortunately though, as Mr-Fox says, SD is so unpredictable it could be anything (even though your post excludes most of the usual options), so I won't even try to guess!

You are correct.
The differences are due to "intentional acts" to put it like that. It's something done/used differently with the intention of improving results, i obviously wasn't sure what it would do as things aren't all that predictable for "us".
I know it's a pretty impossible task and i wasn't expecting answers in that regard, more to make people just go "hmm..." and have a bit of a think about it as the answer(s) themselves isn't all that complex, but the path to finding them probably is.

So for those wanting to remain unknowing for a bit longer DO NOT open the spoiler

Spoiler: Secret information inside, open at great risk

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

 

[Mr-Fox]

You are correct.
The differences are due to "intentional acts" to put it like that. It's something done/used differently with the intention of improving results, i obviously wasn't sure what it would do as things aren't all that predictable for "us".
I know it's a pretty impossible task and i wasn't expecting answers in that regard, more to make people just go "hmm..." and have a bit of a think about it as the answer(s) themselves isn't all that complex, but the path to finding them probably is.

So for those wanting to remain unknowing for a bit longer DO NOT open the spoiler

Spoiler: Secret information inside, open at great risk

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

Oh yeah... I completely forgot about control images and such. I never tried it since many seemed to have mixed results and I wanted to avoid complications.

 

[Mr-Fox]

Your hot boss..

 

[me3]

Something i forgot to post earlier today.
Going back to the sampler posts not too long ago, just some quick details.

• Samplers with a "a" in their name are "versions" that won't converge, and as such you get images which won't be reproduced due to the slight randomness they add
• Going by a measurable "quality" of images both version of DPM++ SDE and DDIM gives the highest quality and can do that at very few steps. Apparently at just 10-15 steps. Downside is that the SDE samplers doesn't converge
• Losing some image quality, but gaining the "ability" to converge, DPM++ 2M karras is seemingly your best option, but need twice the steps as the ones in the point above

Someone said they liked Heun, apparently that's an "improved" version of euler, basically doing two passes instead of one, so it's twice as slow, but seemingly it doesn't produce that much higher quality for most steps. I've not tested this and it's purely a "measured" quality not visually interpreted by peoples eyes, nor does it account for peoples personal preference or artistic "styles".

Really wish i'd learned about this BEFORE i started generating nearly 1000 regularisation images using DDIM at 50 steps when seemingly it would give almost the same quality at 8 steps...I'm putting off testing that as it'd likely piss me off finding out how much time was wasted because of just blindly following peoples "highly recommended" instructions without bothering to do testing for confirmation first.