- Jun 10, 2019
- 223
- 4,376
- Jul 5, 2020
- 14
- 241
- Jun 10, 2019
- 223
- 4,376
- Jan 9, 2019
- 123
- 465
Havent checked the thread in a while, but some of you are excellent.
New update means new version or attempt at something better, I have learned how character tags actually work, in version 3, you can just use name and they will show up, no physical or hair description necessary,
Training is currently 15/33 done.
some info about training, 423 is the amount of pics selected, so less than before I think but less similiar ones, the new updates were included but focus, no point in using variation of same things unless its different enough to teach the model, major difference is the Resolution for training, the training is being done full res, so no resize but 1920X1080, which means the details should be better, previous ones were auto resized to default 1024 maximum dimension.
The faces are coming, I guess, the learning is done in a cosine wave, so rate of learning aren't equal across the board, so this could be the final-ish face or this could be only half or barely learned.
Prompt: score_9, score_8_up, score_7_up, uhd, 4k, highres, masterpiece, beautiful, detailed, rating_questionable, 1girl, Lena
Negative:score_6, score_5, score_4, source_pony, bad quality, lowres, bad anatomy, bad hands, signature, watermarks, ugly, imperfect eyes, skewed eyes, unnatural face, unnatural body, error, extra limb, missing limbs, painting by bad-artist, 3d, ugly
this is my new template I guess of negatives, its not necessary, you can check if it gives better, depending on the pony checkpoint, some are trained to not have negatives and will give you worse results.
As you can see my only real prompts were 1girl and Lena, everything else is just quality prompts, trying to get good result.
Since I first made the Lora, there has been an abundance of Pony checkpoints with their own style and advantage so you pick your favorite
There are also lots of Pony realism checkpoints out there, below was done by my own Mix, so Example same prompts only name was changed, well realism isnt the point, but with Pony it does work if you get a realistic version of it. Its interesting how it works, it wasnt possible before it, just making lora from 2d art and get good realisic results.
Few more examples to end things.
Prompt: score_9, score_8_up, score_7_up, uhd, 4k, highres, masterpiece, beautiful, detailed, rating_explicit, 1girl, breasts, penis, 1boy, nude, nipples, fellatio, oral, all fours, Ivy
And lastly
Prompt: score_9, score_8_up, score_7_up, uhd, 4k, highres, masterpiece, beautiful, detailed, rating_questionable, 2girls, Emma and Ivy, posing
I wasnt sure it will work, because I wasnt sure it would understand when there are 2 different girl tags in 1 picture, but as you can see it kinda works, well this was 3rd attemp, 1st one it merged 2 girls into 1 girl that looks like emma but had Ivy's hair, And 2nd one both looked like Emma, the strong Bias for Emma is because Emma has more pictures of her in the training Data.
Next post I will post the link when training is done and hopefully its good.
New update means new version or attempt at something better, I have learned how character tags actually work, in version 3, you can just use name and they will show up, no physical or hair description necessary,
Training is currently 15/33 done.
some info about training, 423 is the amount of pics selected, so less than before I think but less similiar ones, the new updates were included but focus, no point in using variation of same things unless its different enough to teach the model, major difference is the Resolution for training, the training is being done full res, so no resize but 1920X1080, which means the details should be better, previous ones were auto resized to default 1024 maximum dimension.
The faces are coming, I guess, the learning is done in a cosine wave, so rate of learning aren't equal across the board, so this could be the final-ish face or this could be only half or barely learned.
Prompt: score_9, score_8_up, score_7_up, uhd, 4k, highres, masterpiece, beautiful, detailed, rating_questionable, 1girl, Lena
Negative:score_6, score_5, score_4, source_pony, bad quality, lowres, bad anatomy, bad hands, signature, watermarks, ugly, imperfect eyes, skewed eyes, unnatural face, unnatural body, error, extra limb, missing limbs, painting by bad-artist, 3d, ugly
this is my new template I guess of negatives, its not necessary, you can check if it gives better, depending on the pony checkpoint, some are trained to not have negatives and will give you worse results.
As you can see my only real prompts were 1girl and Lena, everything else is just quality prompts, trying to get good result.
Since I first made the Lora, there has been an abundance of Pony checkpoints with their own style and advantage so you pick your favorite
There are also lots of Pony realism checkpoints out there, below was done by my own Mix, so Example same prompts only name was changed, well realism isnt the point, but with Pony it does work if you get a realistic version of it. Its interesting how it works, it wasnt possible before it, just making lora from 2d art and get good realisic results.
Few more examples to end things.
Prompt: score_9, score_8_up, score_7_up, uhd, 4k, highres, masterpiece, beautiful, detailed, rating_explicit, 1girl, breasts, penis, 1boy, nude, nipples, fellatio, oral, all fours, Ivy
And lastly
Prompt: score_9, score_8_up, score_7_up, uhd, 4k, highres, masterpiece, beautiful, detailed, rating_questionable, 2girls, Emma and Ivy, posing
I wasnt sure it will work, because I wasnt sure it would understand when there are 2 different girl tags in 1 picture, but as you can see it kinda works, well this was 3rd attemp, 1st one it merged 2 girls into 1 girl that looks like emma but had Ivy's hair, And 2nd one both looked like Emma, the strong Bias for Emma is because Emma has more pictures of her in the training Data.
Next post I will post the link when training is done and hopefully its good.
Last edited:
- Jan 9, 2019
- 123
- 465
While writing 3 more steps are done so why not show you the learning rate, as you can see chracter tag working well.
Prompt: score_9, score_8_up, score_7_up, uhd, 4k, highres, masterpiece, beautiful, detailed, rating_questionable, Ivy, screenshot, 1girl, phone, instagram post
Prompt: score_9, score_8_up, score_7_up, uhd, 4k, highres, masterpiece, beautiful, detailed, rating_questionable, Ivy, twerk, ass focus
Also wanted to show some artistic checkpoint like DucHaiten pony, it really makes pretty work
This one came out amazing too
Prompt: score_9, score_8_up, score_7_up, uhd, 4k, highres, masterpiece, beautiful, detailed, rating_questionable, toes, from behind, jess, swimsuit, open clothes, arm tattoo, sitting
1 more realistic Minnerva , cant ignore her.
Prompt: score_9, score_8_up, score_7_up, uhd, 4k, highres, masterpiece, beautiful, detailed, rating_questionable, Ivy, screenshot, 1girl, phone, instagram post
Prompt: score_9, score_8_up, score_7_up, uhd, 4k, highres, masterpiece, beautiful, detailed, rating_questionable, Ivy, twerk, ass focus
Also wanted to show some artistic checkpoint like DucHaiten pony, it really makes pretty work
This one came out amazing too
Prompt: score_9, score_8_up, score_7_up, uhd, 4k, highres, masterpiece, beautiful, detailed, rating_questionable, toes, from behind, jess, swimsuit, open clothes, arm tattoo, sitting
1 more realistic Minnerva , cant ignore her.
Last edited:
Love that Minerva, looks like a brunette Amy Ryan!
- Jan 9, 2019
- 123
- 465
You must be registered to see the links
You guys can download / try V3 for yourself here
I will request that you guys do some testing on the generation and quality, because there were 33 Epochs, it could be a little over trained, I have all the epochs saved but today I dont have the time to test each one to see at what point training was unnecessary. Should be fine.
Last edited:
- Jun 10, 2019
- 223
- 4,376
You must be registered to see the links
so i found best reference for Holly x Marcus
- Mar 5, 2019
- 92
- 349
- Mar 5, 2019
- 92
- 349
- Jun 10, 2019
- 223
- 4,376
- Jun 10, 2019
- 223
- 4,376
- May 4, 2017
- 135
- 257
You are usually better off by using control net or faceid to drive the AI towards the character you want. But in general for Pony models and loras stick to a straight forward keyword format like, 1girl, black hair, long hair, straight hair, blue eyes... etc. instead of longer descriptive texts like with 1.5 models. The loras from Tiger Ace are good for mimicking Eva's Artstyle but are pretty neutral when it comes to characters, which is good. ORS has a huge cast and training a lora without bias towards any character is tricky.
Something that works well for me is to generate a bunch of face close ups (different angles) of the character i want first and then use these images with control net or faceid. If your model refuses to give you any resemblance to the original characters even with the lora(s) applied then crop a few faces from the images from ORS and use them as guides. And last but not least, inpaint. Inpaint whatever gets screwed up by the AI, which is usually the eyes or mouth area. But you can also use something like FaceDetailer from the Impact or InspirePack (ComfyUI) or the
You must be registered to see the links
toolsets.
- Jun 10, 2019
- 223
- 4,376
Well... I made some stuff with TIGERs Lora and i have to say its amazing, i have like 3 days of experience in stable diffusion and i came up with this. That just shows how powerfull is ai nowdays and it will be better and better. I used some diffrent styles, and checkpoints. I had a lot of fun with this 
Cindy
Holly
Lena
Cindy
Holly
Lena
- Jun 10, 2019
- 223
- 4,376
