- May 4, 2019
- 456
- 1,635
I wanted to reproduce the pictures from an earlier post I did in this thread from 5th January, and it proved to be almost impossible! So I just had to produce a bunch of similar images using the same prompt instead:
There were actually a bunch of changes to the webui project so that the model hashes changed, and also changes to how the samplers work. In the newest versions there are now settings to switch back to the old behaviour which helps, but still I just couldn't figure out what I was missing anymore.
I did find that reproduceability is not as easy that I thought in general, I tried to reproduce a picture Sepheyer posted here, but I just couldn't get the same result as him:
It looks very similar, but something must have still been different.
So, instead of trying to figure out why it didn't work, I decided my time is better spend on trying something else. So I trained a LORA with the pictures of that Daz character I already used as example in my first post here. Basicly I trained with the pictures of this little series:
And then I tried applying that LORA to Sepheyer's prompt:
And I was actually really surprised how easy training a LORA was and how relatively well it worked. In particular the hair and the clothing was picked up (only the shirt, the skirt was barely present in the training data, still came out looking like I expected it though), even tough the training data wasn't really that comprehensive. Though I think the glasses were not a good thing to try and use for a first attempt at training a LORA. They caused ugly effects when applying the LORA with too much weight, and also it was kind of difficult to then make it create an image without glasses.
I also applied the LORA to pictures of the training data, with inpainting. It kind of worked as a way to recreate the same character, but with a certain style. More realistic, or more anime like, some before after pics:
There were actually a bunch of changes to the webui project so that the model hashes changed, and also changes to how the samplers work. In the newest versions there are now settings to switch back to the old behaviour which helps, but still I just couldn't figure out what I was missing anymore.
I did find that reproduceability is not as easy that I thought in general, I tried to reproduce a picture Sepheyer posted here, but I just couldn't get the same result as him:
It looks very similar, but something must have still been different.
So, instead of trying to figure out why it didn't work, I decided my time is better spend on trying something else. So I trained a LORA with the pictures of that Daz character I already used as example in my first post here. Basicly I trained with the pictures of this little series:
You must be registered to see the links
And then I tried applying that LORA to Sepheyer's prompt:
And I was actually really surprised how easy training a LORA was and how relatively well it worked. In particular the hair and the clothing was picked up (only the shirt, the skirt was barely present in the training data, still came out looking like I expected it though), even tough the training data wasn't really that comprehensive. Though I think the glasses were not a good thing to try and use for a first attempt at training a LORA. They caused ugly effects when applying the LORA with too much weight, and also it was kind of difficult to then make it create an image without glasses.
I also applied the LORA to pictures of the training data, with inpainting. It kind of worked as a way to recreate the same character, but with a certain style. More realistic, or more anime like, some before after pics:
Last edited:
- Jan 24, 2020
- 1,401
- 3,808
How cute, and absolutely wholesome in this otherwise perverted thread..
I do love me some boobas and badonkas, but I really wouldn't mind more content like this.
I have thought of doing something not perverted also but I keep getting distracted some how..
- Jan 24, 2020
- 1,401
- 3,808
Awsome job. If you wouldn't mind. Could you make a post in the learning thread about making lora's?
I understand that you just tried it, but you did learn a lot I bet. Many people me included are guaranteed to be curious about trying it ourselves.
- May 4, 2019
- 456
- 1,635
Actually going to do that in a bit. It was really easy using this:
You must be registered to see the links
I am going to upload my training data as an example together with the trained LORA.
- Oct 29, 2017
- 812
- 2,200
- May 18, 2018
- 215
- 2,167
- Dec 21, 2020
- 1,611
- 3,863
This might be useful, or completely not. See how the girl on the left has grain in her face and hair? These are actually artefacts of the face restoration algorithm. There are two major algos for "face restore". I assume you are using WebUI by A1111, and they use CodeFormer as the built-in option. Lots of times it works well and seamlessly, tho sometimes it has these artefacts that one discards assuming this is grain and a pleasant addition to the image.
So, just so the folks know there is an alternative algo, GFPGAN. And here is how these algos compare: link.
A1111 has an option to use GFPGAN - this can get rid of that "staircase / grain" artefact in the woman's hair.
This is what GodOfPandas's image looks like with artifacts removed by GFPGAN:
"How to" here: link.
Last edited:
- Jan 10, 2019
- 323
- 1,749
His tutorial videos are super easy to follow and great for learning. Meantime in the mountains, no naked bird but a rugged mountain man:
- Oct 10, 2017
- 7,628
- 21,959
Look, i'll be honest. I really don't care about the bird there but that robot in your sig is seriously adorable!
- Oct 29, 2017
- 812
- 2,200
Stable Diffusion and I read advice on prompts to have a better quality. And sometimes Photoshop to improve the picture.
