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.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..
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".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.
), same seed/step/cfg/prompt etc. 
I love this kind of experimentation. Looking forward to the next update. Good work.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
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'm reading me3's post differently - I think it's a challenge for us, not a request for some kind of information or assistance.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.
You are correct.
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.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
You don't have permission to view the spoiler content. Log in or register now.
Try downloading the Textual Inversion called badhands4 or badhands5 and it should help with that. I don't know if I can post links or not but you can find it on Civitai. I'm new to all this and I'm figuring it out as I go lol someone told me about badhands.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.
I've noticed that deebooru also gives better prompts for NSFW art.
Yes you can post links, no problem. If you look around in this thread we do it a lot. I'm familiar with "badhands" however Ti's has been very problematic in my experience so I choose to stay away from them for now.
A bit surprised you've been having problems with TIs given the amount of time i've spent with them the only similar "issue" i've experience is that if you use one in a prompt then delete or replace the TI file it doesn't get updated and you have to either relaunch or hit the refresh for embeddings.
)
Yeah you could be correct about needing to do the reload embedings. When I had the issue it would stay even after restarting SD completely, maybe even after PC reboot, that part I don't remember. It only disappeared after many many generated images. It was like I had to purge the "memory" by doing many new generations. I have had the same happen with hypernetworks, so after this I stay well clear of those "things"..A bit surprised you've been having problems with TIs given the amount of time i've spent with them the only similar "issue" i've experience is that if you use one in a prompt then delete or replace the TI file it doesn't get updated and you have to either relaunch or hit the refresh for embeddings.
I've never had it gotten "cached" in the actual prompts though, soon as they've been removed from there the effects have gone as well.
However I do remember having some oddities with just pure prompts. On some very few cases i've added something to the prompt, generated a few sets of images, noticed that it seemingly didn't have any effect, removed it again then only to have it show in the next couple of images.
It's almost like the AI seems to "remember" certain things too. Like things you use very often in your prompts showing up on their own or the "style/quality" you always use seem to be applied. Because people mostly just use the same prompts every time, specially negative ones, we don't notice or just overlook it as we're used to seeing it.
(They are planing to take over the world and wipe us all out, so obviously they need to know what we prefer to easily manipulate us
I have no idea why or how this would happen and so until I have a clear answer and either know why or how to avoid it/remove it I will stick with Loras.
I remember seeing someone post a set of images with bodypaint. If it's good or not is up to you to decide.
