Turns out PonySDXL can’t do armpit sweat pit stains or any sweat on shirt properly either. This will fix it. It’s trained towards armpit stains, with less focus on ?????, belly and back.
It had a lot of women on the trained data, it works regardless of gender.
This is a Pony version of my other LoRAs. I used the same process here.
Please use each version accordingly. The anime was epoch 20; realistic was epoch 18. Not only that, but their blocks analysis were very different, that is why I'll release two versions.
If your base checkpoint for generation is a realistic merge with Pony, not PonyXL 6, my other base SDXL LoRA might work better than this realistic Pony version.
Recommended weight is 1.
Good from 0.8 up to 1.5 or further
This LoRA works at high weights. That is why I rescaled it. It has a low effect at 0.8, medium at 1.0 and high at 1.15. But you can go to 1.5 or more.
I only used trigger keywords to see if I could control the stains’ location (not very effective), it doesn’t need them! They were not the first tag on the caption, but it helps to use.
Also used neg and pos embeddings (xxx or safe).
Trigger Keyword:
s3t on armpit, s3t on ?????, s3t on belly, s3t on back
Supporting prompts:
score_9, score_8_up, score_7_up, (shirt with sweat:1.3), wearing clothes, sweat-spots-smudges
Negative supporting prompts:
score_3_up, score_4_up, score_5_up, black dye, black paint, black smudges, blue paint, yellow paint, worst quality, jpeg artifacts, low-res
I strongly recommend for better results to use any of the @prgfrg23 Styles for Pony:
Smooth 2 for anime
Photo 2 for realistic
You can combine this with Loractl to make sure you’ll get your composition. Set the Lora to start with low weight at 0.33%, increase it after.
<LoraName:[email protected],[email protected]>
The base checkpoint is the “ponyDiffusionV6XL”. Works with pony based SDXL. I did not use quality tags for training but you should use them on the generations.
I'm not well versed on Pony. For anime I recommend Pony itself.
For realistic: everclearPNYByZovya_v2VAE OR Pony base + LoRa Style Photo 2 <Style SDXL_LoRA_Pony Diffusion V6 XL:0.7>.
Do not use Photo 2 with everclearPNY.
Also, so far in my testings, for merges like damnPonyxlRealistic_damnV10 my standard base SDXL Sweaty Shirt works better than this realistic Pony version.
Turbo Pony worked as well.
For standard generation
HIGHER CFG works better for Pony: 7.0
Euler a
Problems with the current Lora:
Sometimes they might “????” themselves. Or, well, groin sweat. That’s a benefit in my book.
Underwear sweat is not well trained, but might work.
Advanced training settings:
Anime version: Remerged with these block weights: lbw=0,0,0,0,0.8,0,1,1,1,1,1,1 + (0.1) of my Standard Sweaty_Shirt_SDXL. Also a re-scale from 1.2 to 0.8 using Ostris tool.
Realistic version: Remerged with these block weights: lbw=0.125,0,0,0,0.2,0,0.4,1,1,0,1,1 + (0.1) of my Standard Sweaty_Shirt_SDXL. Also a re-scale from 1.2 to 0.8 using Ostris tool.
The anime version and the realistic (soon) version were all trained on the same dataset (realistic images). But the anime version came out quite good even with no anime on training.
Some settings: Trained 1024 res. 190 images. Epoch 20 from 44. now prodigy 1.0. 2 steps folder. constant BATCH 6, rank 8/1,Scale weight norms 1, snr gamma 5, Noise offset 0.0357, no regularization image, Max Token Length 225
Hopefully you can leave some results and some comments. Any idea is appreciated. Thank you.
