[Stable Diffusion] Prompt Sharing and Learning Thread

[modine2021]

No probs, this is the way I do it - there must be better ways, but it works for what I do.

1. select a jpg
2. add it to your openpose if that's what you want and detect from image as per this screenshot: View attachment 2511457
   
3. once finished, I just save it as png somewhere on my PC
4. I go to txt to image and then I open control net, import the png I just created and select 'openpose' as per this: View attachment 2511459 I don't need to select a 'processor' as I've already done the work in openpose.
5. type your prompts and generate View attachment 2511462
   
6. get these: View attachment 2511463 View attachment 2511464 View attachment 2511465

The other way I use is to just get to txt to img:

1. View attachment 2511466 attach the image you want to work with directly in the controlnet section, select your preprocessor and model. With this image and 'depth' I get these results: View attachment 2511468 View attachment 2511469 View attachment 2511471 using these prompts: View attachment 2511473
   
2. You can change the preprocessor and model, use same prompts and you get different results, here with 'hed' instead View attachment 2511476 View attachment 2511477 View attachment 2511478

Hope this helps. Others may have more tips and tricks, but this works fine for me at the moment.
I use realistic vision 2.0 (

You must be registered to see the links

) as a model.

finally got it acting right. took 50/11 times. but it now remembers the pose from the image. still need a bit practice( the feet wrong in this one)

original pic




OpenPose pic

 

[sharlotte]

Yeah, practice makes perfect! Vary the prompts, the preprocessor types and models and you will get different results, it's all a steep learning curve.

 

[devilkkw]

have some of you nice suggestion on Textuail inversion?
i experimented a lot on these day, and i see many tutorial, but every tutorial use same "learning rate", and so i experimented with it.

On my test seem work great on high value, i mean i train on 1200 step and get good result with value like "1.05, 3.2,2.1"
Have someone experienced it?

 

[sharlotte]

I read in different TI on CIVITAI people using 0.004 or 0.005 and AI Entrepreneur uses the following '0.05:10,0.02:20,0.01:60,0.005:200,0.002:500,0.001:3000,0.005' - I have not yet successfully completed on. Just starting it and seeing over 4-5 hours of training was a bit deflating...

 

[Mr-Fox]

have some of you nice suggestion on Textuail inversion?
i experimented a lot on these day, and i see many tutorial, but every tutorial use same "learning rate", and so i experimented with it.

On my test seem work great on high value, i mean i train on 1200 step and get good result with value like "1.05, 3.2,2.1"
Have someone experienced it?

I have not tried out making my own TI. When I tried TI's by other's they only created problems for me such as taking over too much and breaking the prompt etc, even at low weight. And after I removed them they tend to linger.
So atm they are cancer in my view. But disregarding all that, when I made my Lora I found that slow learning was the ticket.
I have no idea if it transfers to TI's.

 

[devilkkw]

I read in different TI on CIVITAI people using 0.004 or 0.005 and AI Entrepreneur uses the following '0.05:10,0.02:20,0.01:60,0.005:200,0.002:500,0.001:3000,0.005' - I have not yet successfully completed on. Just starting it and seeing over 4-5 hours of training was a bit deflating...

Wow, is high time. at 3000 step my train run in 30 mins.
Then in 1200 step i run it in 14 mins.
Can you prepare your image for train?

I have not tried out making my own TI. When I tried TI's by other's they only created problems for me such as taking over too much and breaking the prompt etc, even at low weight. And after I removed them they tend to linger.
So atm they are cancer in my view. But disregarding all that, when I made my Lora I found that slow learning was the ticket.
I have no idea if it transfers to TI's.

Yes, i decide to trin my own because many on downloaded are really messed up, seem bad initialization word or too many step.
I'm testing now actually i reach good result from 3 to 14 cfg.
Also is not powered like LORA, is possible to training lora on 6Gb?

 

[Mr-Fox]

Here's some rentry's I found on this subject.

You must be registered to see the links

You must be registered to see the links

You must be registered to see the links

 

[Mr-Fox]

Wow, is high time. at 3000 step my train run in 30 mins.
Then in 1200 step i run it in 14 mins.
Can you prepare your image for train?

Yes, i decide to trin my own because many on downloaded are really messed up, seem bad initialization word or too many step.
I'm testing now actually i reach good result from 3 to 14 cfg.
Also is not powered like LORA, is possible to training lora on 6Gb?

I think that you can make your own Lora with 6GB vram yes. You just need to use low resolution 512,512 or maybe you need to go lower.
Rentry's I recommend to read if you decide to try it.

You must be registered to see the links

You must be registered to see the links

I used kohya_ss to make mine.

 

[Mr-Fox]

I read in different TI on CIVITAI people using 0.004 or 0.005 and AI Entrepreneur uses the following '0.05:10,0.02:20,0.01:60,0.005:200,0.002:500,0.001:3000,0.005' - I have not yet successfully completed on. Just starting it and seeing over 4-5 hours of training was a bit deflating...

4-5 hours isn't bad at all. Try 10 hours plus for a good Lora... At least with my setup. Just set it and go do some laundry or something..

 

[Jimwalrus]

have some of you nice suggestion on Textuail inversion?
i experimented a lot on these day, and i see many tutorial, but every tutorial use same "learning rate", and so i experimented with it.

On my test seem work great on high value, i mean i train on 1200 step and get good result with value like "1.05, 3.2,2.1"
Have someone experienced it?

I have made a few TIs (downloadable from

You must be registered to see the links

) - some people run them high learning rate to start with, then turn it down massively as it starts to produce good results. Frankly they always look a bit off or rushed to me.
I have a fairly standard 'recipe' for a photorealistic-type TI:

- At least 60 images. Good ones. Varied poses, clothes and, especially, expressions. Try to get a face in profile or two. Some full body, some close-up faces. Curate them well and ensure you manually crop and resize them to 512x512. Don't use automated tools, they're all shit! MS Paint is fine. Ideally, have over 100 images. Don't upscale shit photos, go looking for a better quality original instead. All this takes time. It matters though. Don't bake a cake with mouldy flour and rotten eggs.

- Some guides recommend the mirroring option to double the number of images. This is fine if you're training on an object, but how many people have a truly symmetrical face? You'll just make them look a bit weird. Don't use this option.

- Use a completely unambiguous name and token for your embedding e.g. for Kylie Minogue I used "KYL13M". That way SD won't try to use its default Kylie training which looks nothing like her. No chance of any influence.

- A learning rate of 0.0015. I don't vary it, seems OK for me. It's pretty low so you shouldn't overtrain. Overtraining is mostly an issue for the impatient who try to rush it.

- 12-20 Vectors per token. No less than 12 for a person, ensure you have at least 5 images per vector. Don't bother with more than 20. I once tried 40, the law of diminishing returns kicked in.

- Use SD1.5 as the base model for training to give optimum applicability across other models.

- Set 'subject' instead of 'style'.

- Set a good prompt in the txt2img tab - one you would probably use when generating the finished product. Avoid weightings though, keep it all unweighted. Avoid LoRAs or other embeddings too if you can. Set it to preview (and save an embedding) every 500 or 1000 steps. Give it a fixed seed so you're always generating previews on the same prompt and seed.

- Set the steps limit to 30,000.

- Hit 'Train embedding', open the preview images folder and find something else to do for several hours(!)

- Your previews will be terrible, SD1.5 is actually an awful model for photorealistic people. You should still be able to find the point where it really starts to settle down though, usually around 17k steps or more.

- Copy the relevant embeddings (the .pt files, suffixed "-n" where n is the no of steps) for those previews that seem best into the main \embeddings folder.

- Set up scripting X/Y/Z plot with prompt S/R to switch from KYL13M-18000 to KYL13M-19000 to KYL13M-20000 etc. Run a quick test in your model of choice NOT SD1.5 itself. A couple of images for each draft embedding on test, with a fixed seed so it reuses the same couple of seeds each time.

- Compare the results to see which is best. Remember, if you've spent long enough curating the images, the not-Artificial Intelligence in your brain will also be trained on the subject! You should inherebtly know which are closest to the real deal.

- Run some more test images to check. Try a few really wacky prompts (including nudity) to ensure you've not baked in clothes, poses, expressions or background items.

- Run some good SFW images off as previews for Civitai and upload it there for everyone to download.

 

[devilkkw]

This is good. thank you. i will try it.

I think that you can make your own Lora with 6GB vram yes. You just need to use low resolution 512,512 or maybe you need to go lower.
Rentry's I recommend to read if you decide to try it.

You must be registered to see the links

You must be registered to see the links

I used kohya_ss to make mine.

Thank you for link, inspect it and try.

 

[Jimwalrus]

I have made a few TIs (downloadable from

You must be registered to see the links

) - some people run them high learning rate to start with, then turn it down massively as it starts to produce good results. Frankly they always look a bit off or rushed to me.
I have a fairly standard 'recipe' for a photorealistic-type TI:

- At least 60 images. Good ones. Varied poses, clothes and, especially, expressions. Try to get a face in profile or two. Some full body, some close-up faces. Curate them well and ensure you manually crop and resize them to 512x512. Don't use automated tools, they're all shit! MS Paint is fine. Ideally, have over 100 images. Don't upscale shit photos, go looking for a better quality original instead. All this takes time. It matters though. Don't bake a cake with mouldy flour and rotten eggs.

- Some guides recommend the mirroring option to double the number of images. This is fine if you're training on an object, but how many people have a truly symmetrical face? You'll just make them look a bit weird. Don't use this option.

- Use a completely unambiguous name and token for your embedding e.g. for Kylie Minogue I used "KYL13M". That way SD won't try to use its default Kylie training which looks nothing like her. No chance of any influence.

- A learning rate of 0.0015. I don't vary it, seems OK for me. It's pretty low so you shouldn't overtrain. Overtraining is mostly an issue for the impatient who try to rush it.

- 12-20 Vectors per token. No less than 12 for a person, ensure you have at least 5 images per vector. Don't bother with more than 20. I once tried 40, the law of diminishing returns kicked in.

- Use SD1.5 as the base model for training to give optimum applicability across other models.

- Set 'subject' instead of 'style'.

- Set a good prompt in the txt2img tab - one you would probably use when generating the finished product. Avoid weightings though, keep it all unweighted. Avoid LoRAs or other embeddings too if you can. Set it to preview (and save an embedding) every 500 or 1000 steps. Give it a fixed seed so you're always generating previews on the same prompt and seed.

- Set the steps limit to 30,000.

- Hit 'Train embedding', open the preview images folder and find something else to do for several hours(!)

- Your previews will be terrible, SD1.5 is actually an awful model for photorealistic people. You should still be able to find the point where it really starts to settle down though, usually around 17k steps or more.

- Copy the relevant embeddings (the .pt files, suffixed "-n" where n is the no of steps) for those previews that seem best into the main \embeddings folder.

- Set up scripting X/Y/Z plot with prompt S/R to switch from KYL13M-18000 to KYL13M-19000 to KYL13M-20000 etc. Run a quick test in your model of choice NOT SD1.5 itself. A couple of images for each draft embedding on test, with a fixed seed so it reuses the same couple of seeds each time.

- Compare the results to see which is best. Remember, if you've spent long enough curating the images, the not-Artificial Intelligence in your brain will also be trained on the subject! You should inherebtly know which are closest to the real deal.

- Run some more test images to check. Try a few really wacky prompts (including nudity) to ensure you've not baked in clothes, poses, expressions or background items.

- Run some good SFW images off as previews for Civitai and upload it there for everyone to download.

BTW this was intended more as a 'TIs for Dummies' and/or a brain-dump of my workflow rather than any attempt to patronise you devilkkw, you've obviously made some of your own so know much of this already.

As you can probably see though, to me the quality of the result is as much from the preparation of the training data and the QC checking of the results than the actual training itself.

 

[devilkkw]

Is just try to be better, for some model i get god result with setting i posted, but as i look at your seem much better,so i'm testing it now.
Also preparing image is a time spending, but if you want good result you have to do it.
I have ready set so i'm triyng your setting now. 3 hours i think is good for 30000 step.

 

[Jimwalrus]

Is just try to be better, for some model i get god result with setting i posted, but as i look at your seem much better,so i'm testing it now.
Also preparing image is a time spending, but if you want good result you have to do it.
I have ready set so i'm triyng your setting now. 3 hours i think is good for 30000 step.

Please, please let me know your results - I am no expert, my 'recipe' is just what has worked for me so far. I don't know if there's something over-trained in my embeddings or that otherwise causes issues. Any and all feedback is greatly appreciated, even if it's just that you've found nothing weird and it produces flexible models.

 

[devilkkw]

after testing i give you result. i test it in different models after finishing train.

 

[devilkkw]

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.

 

[Jimwalrus]

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.

Thanks a lot for this, I'll be giving some or all of these a try.
The 30k steps in my original 'recipe' was only intended as a backstop, I generally find best results with my settings at around 18k-25k, usually ~21k.

 

[devilkkw]

i saved all train at 1000 step, the 19k give me good result, but the problem is the same.
Maybe is possible problem come with models used for test, on standard sd1.5 works great, but as you know not much good result in composition.
Let me know your impression if you test my setting. thank you.

 

[Sepheyer]

Anyone tried erotic story generators? Any recommendations?

I tried

You must be registered to see the links

Spicy.

Let's see if getting AI to write erotic stories generates enough interest to start its own thread.

Spoiler: AI Generated Adult-ish Story

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

Spoiler: Prompt

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

 

[Elefy]

AI to write erotic stories generates enough interest

... and its time to generate scene images for the story