[Stable Diffusion] Prompt Sharing and Learning Thread

[devilkkw]

All useful post, now i've understand. Also made my custum impaint model.
Most important is when you merge, don't make any yaml file, because if you make it, model is not loaded correctly and get error about tensor size.

Also, after done process linked by miaouxtoo, i had better result doing another step with new inpainting model:
do a weightsum at 1 with model you used for obtaining inpainted one.

 

[daddyCzapo]

Hi! I recently started messing around with sd, and wanted to share with you some of my very basic prompts, at first i just fulfilled my perv fantasy, so bare with me

Spoiler: 1

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

Spoiler: 2

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

Spoiler: 3

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

Spoiler: 4

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

Most of them have vary basic prompts, and are fucked up in some kind of way.. but i like them and just wanted to share with you guys. Models used are either experience or pfg, don't know if that's in the png data

 

[Mr-Fox]

Hi! I recently started messing around with sd, and wanted to share with you some of my very basic prompts, at first i just fulfilled my perv fantasy, so bare with me

Spoiler: 1

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

Spoiler: 2

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

Spoiler: 3

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

Spoiler: 4

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

Most of them have vary basic prompts, and are fucked up in some kind of way.. but i like them and just wanted to share with you guys. Models used are either experience or pfg, don't know if that's in the png data

This is very good and better than most that just started. Yes the checkpoint(model) is always in the data.
For the realistic stuff I would like to recommend that you try out

You must be registered to see the links

.
devilkkw have created merged checkpoints that might be good for semirealistic and stylized stuff, at the very least worth checking out.
His latest:

You must be registered to see the links

I'm looking forward to see more from you.

 

[daddyCzapo]

This is very good and better than most that just started. Yes the checkpoint(model) is always in the data.
For the realistic stuff I would like to recommend that you try out

You must be registered to see the links

.
devilkkw have created merged checkpoints that might be good for semirealistic and stylized stuff, at the very least worth checking out.
His latest:

You must be registered to see the links

I'm looking forward to see more from you.

Oh, i will definitely try these out, thanks. I tried to do as much research beforehand as i can so it's not completely garbo From my experience, experience_70 is great for the "wet, sweaty, shiny" look that i was going for, and pfg gives more cartoon-ish vibe to the generated images.
And here's one more, that i forgot to add in the first post:

Spoiler: 1

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

 

[Mr-Fox]

I have been working on my first Lora. I had a lot of problems at first and had to do a lot of trial and error.
However now I'm starting to get better results.
If you have a pascal 10series card like me, you might get an error and can't use AdamW8bit. I tried to use Lion as recommended but got terrible results. I have set it to get a sample image last step for each source image so I can follow the process.
I found a very good guide

You must be registered to see the links

. This article gave a solution.
You need to download "

You must be registered to see the links

" rename the file libbitsandbytes_cudaall.dll to libbitsandbytes_cuda116.dll, put it in sd-scripts\svenv\Lib\site-packages\bitsandbytes\ (overwrite). Download the 0.35 one

You must be registered to see the links

.
Since I use kohya ss, I put mine in kohya_ss\venv\Lib\site-packages\bitsandbytes.
Now I can use AdamW8bit and is getting much better results. The guides videos I have watched recommended to use clip 2.
I will try to use clip 1 and see if it makes a difference. An observation is that the checkpoint you are using for your training could be something very important that makes a big difference. I don't know yet wich one is a good choice.
To be continued..

 

[Mr-Fox]

Oh, i will definitely try these out, thanks. I tried to do as much research beforehand as i can so it's not completely garbo From my experience, experience_70 is great for the "wet, sweaty, shiny" look that i was going for, and pfg gives more cartoon-ish vibe to the generated images.
And here's one more, that i forgot to add in the first post:

Spoiler: 1

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

Instead of wet skin, wich get you the pink slime. I'm assuming here because I can't run SD to check the png info while I'm doing Lora training. Try to use moist skin or oiled skin or oily skin etc.
Liquids are problematic for SD as of yet. Oh and it's well worth the cost to upgrade your pc to 32Gb of ram if you can.

 

[daddyCzapo]

Instead of wet skin, wich get you the pink slime. I'm assuming here because I can't run SD to check the png info while I'm doing Lora training. Try to use moist skin or oiled skin or oily skin etc.
Liquids are problematic for SD as of yet. Oh and it's well worth the cost to upgrade your pc to 32Gb of ram if you can.

actually, i used both, so that's probably why the results are mixed as of right now, a lot of what i generate depends highly on the seed, so i just do the trial and error, by generating 10-20 images in a batch, picking the best looking ones and then using the seed, dialing the details a bit.

 

[Mr-Fox]

A couple of guides by myself I recommend that you read:
https://f95zone.to/threads/stable-diffusion-prompt-sharing-and-learning-thread.146036/post-10138732 (New VAE)
https://f95zone.to/threads/stable-diffusion-prompt-sharing-and-learning-thread.146036/post-10333003 (Large SD Guide)
https://f95zone.to/threads/stable-diffusion-prompt-sharing-and-learning-thread.146036/post-10306143 (cuda memory fix)

 

[devilkkw]

When i merge checkpoint, i do a lot of test before release it.

This is one of some test: how a word change everything.

Prompt: beautiful TYPE portrait of a Spanish woman, details, elegant, centered, symmetrical
Negative prompt: (bad quality, worst quality, low quality:1.2) render

Type= type of image, or something your fantasy think


Action figure

Anime

Crayon

Marble

Pencil

Pixel art

Sand

Simpsons

Splatter


There are many more, just try.

 

[Mr-Fox]

My experimentation with creating lora's is continuing. From my last post I mentioned testing if using clip 2 is good or bad.
Even though I only have a limited amount of data, it's seems so far that it is indeed giving an improvement over not using it.
I found a setting I had completely missed and that no tutorial video is mentioning and it is Noise Offset.
After reading a few posts about it I'm still not sure what it does. I will try it out in the next training session.
You can set a value between 0-1. In the thread I found about it they said 0.2 looks promising.
To be contiued..

 

[Mr-Fox]

Great information sources for lora training.

You must be registered to see the links

(both dreambooth SD extension and kohya)

You must be registered to see the links

(more extensive and kohya focused)

Noise Offset : Increases dynamic range (brighter brights, darker darks). May "deep fry" if set too high.
What this means in practice I have no idea, good to know nonetheless.

 

[devilkkw]

Nice one. please add prompt as rule.

My experimentation with creating lora's is continuing. From my last post I mentioned testing if using clip 2 is good or bad.
Even though I only have a limited amount of data, it's seems so far that it is indeed giving an improvement over not using it.
I found a setting I had completely missed and that no tutorial video is mentioning and it is Noise Offset.
After reading a few posts about it I'm still not sure what it does. I will try it out in the next training session.
You can set a value between 0-1. In the thread I found about it they said 0.2 looks promising.
To be contiued..

On my pc i can't try Lora, but i trained some Textual inversion.
Lora have many setting's to set, hope you fuond best. You are training lora for style or for person-subject?

 

[Mr-Fox]

View attachment 2487565

Nice mix of styles. How did you achieve this? Was it a normal txt2img or did you something "extra" ?

 

[fanciful_dream]

Does anyone know how the | works in webui prompt syntax? Or do you know any good links which explain the prompt format and all the symbols?

 

[fanciful_dream]

Nice mix of styles. How did you achieve this? Was it a normal txt2img or did you something "extra" ?

Text2Img. Using ControlNet with 0.2 weight as well, but usually I don't use it. Also using my personal LoRa for more painting style.

 

[Mr-Fox]

Nice one. please add prompt as rule.




On my pc i can't try Lora, but i trained some Textual inversion.
Lora have many setting's to set, hope you fuond best. You are training lora for style or for person-subject?

I guess a bit of both but technically it's a character,

You must be registered to see the links

.

 

[Mr-Fox]

Text2Img. Using ControlNet with 0.2 weight as well, but usually I don't use it. Also using my personal LoRa for more painting style.
View attachment 2487601

Thank you. ControlNet is next on my list of things to explore.

 

[fanciful_dream]

Thank you. ControlNet is next on my list of things to explore.

It's very easy. I use depth.

 

[fanciful_dream]

You must be registered to see the links

This is interesting

 

[Mr-Fox]

It's very easy. I use depth.

Just one thing. Make sure too share the prompt since this is a prompt sharing thread. You can find the thread rules on page one.
If you don't want to, there is an Ai art thread for showcasing your creations.
ai-art-show-us-your-ai-skill