Sweaty Shirt / Armpit Sweat / Pit stains / Wet Spots – SD 1.5

This is the SD 1.5 version of my SDXL lora. SD can’t do armpit sweat pit stains or any sweat on shirt properly. This will fix it. It’s trained towards armpit stains, with less focus on chest, belly and back.

It had a lot of women on the trained data, it works regardless of gender.

This SD1.5 version took way much more work and to be honest, it's not even close to how good the SDXL version is. So I recommend to use the SDXL one. It's a simple concept (low rank, low size). I did an extensive epoch and block analysis. This one (and most non-pose LoRAs) benefits from a lower Text Encoder (base) weight and zeroing INS and MID blocks.

Recommended weight is 1.

Good from 0.8 up to 1.3 or further

I’m publishing this modified remerged LoRA as the main and not the full block one because people won’t read or cannot apply the best settings and Civitai generations don’t have these settings either way.

So, this is a remerge with these block weights: lbw=0.4,0,0,0,0,0,0,0,1,1,1,1,1,1,0,0,0 and also a re-scale from 1.2 to 0.8 using Ostris tool. 

It should not change your original character or composition even at high weight. But you can lower the TE even further or even do a 0 weight TE if you want lora:name:0.2:1

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, it doesn’t need them! They were not the first tag on the caption, but it helps to use. 

Trigger Keyword:

s3t on armpit, s3t on chest, s3t on belly, s3t on back

Supporting prompts:

Shirt; (shirt with sweat:1.3); wearing clothes; sweat-spots-smudges (can help V5 sometimes)

Negative supporting prompts:

Black dye, black paint, black smudges, blue paint, yellow paint, necklace, worst quality, jpeg artifacts, low-res

DON’T use the word “wet” as this token concept bleeds heavily in any model. You’ll get rain and water everywhere. This is not my LoRA problem, you should just not use this word unless that is what you want.

Before doing this LoRA, I did an extensive word analysis (I did this on SDXL not SD1.5) and the best words for the non-Lora model to do something close to sweat on armpit was the term “sweat-spots-smudges on armpit”. I trained with these trigger words, and after block analysis it performs pretty much the same as the unique token one “s3t”.

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 original “v1-5-pruned.safetensors” 7.7GB. I wonder if it would come out better by training on a realist model... BUT,because I didn't, it should be very flexible with any checkpoint.

I recommend realisticVisionV51 (also the inpainting version), Reliberate_v20, objectiveReality_v20

For standard generation

Lower CFG works better specially is you use the lora at high weight: 4.5

Clip skip: 1. Clip skip 2 can give more consistent clothes format and composition but will mess up colors. It’s worth trying.

DPM++ 2S a Karras

DPM++ 2M Karras

DPM++ SDE Karras

I really strongly recommend to use pertubed-attention, it makes the LoRA ate high strengh way more consisntent.

Obviously, as this is a SD1.5 USE high-res fix and Adetailer without the LoRA.

Kohya deep shrink HRFix is also recommended.

‎

Want to have some “fun”? Install wildcards dynamic prompts extension: https://github.com/adieyal/sd-dynamic-prompts . And wildcard manager. Here is a prompt I made for testing. Paste on prompt:

{side view|front} view of {Harrison Ford|Leonardo DiCaprio|Brad Pitt|Chris Hemsworth|Hugh Jackman} on his {shorts|suit pants|jeans:1)} wearing a {long sleeves|tank top|button|vneck|high crewneck} {light|dark}-__YetAnotherWildcardCollection-main/Colors/Basic__ (shirt with sweat:1.3) {sitting on the floor|running|lying resting on the ground} in a {forest|road|street|distant planet moon surface|marathon} alone, dynamic pose, {his arms are fully extended and he is posing for the camera|his arms are on his hips|his arms are pointing to {his armpits|the camera}|his arms are behind his head}, large (s3t on armpit:1.2), natural lighting, 4k uhd, dslr, high quality, Fujifilm XT3

Problems with the current Lora:

• The worst part about this LoRA is that with tank tops it will sometimes add a black stain on the armpit instead of the proper naked armpit.

• I've seen also some kind of a tendency to show up some kind of "brooch" on the chest.

• "Necklace" is a good word to either add on negative or positive. Not adding it sometimes gives a weird line on the shirt.

• This version is a little weird with wrinkles. I suggest to heavily prompt for a "plain" shirt or simple shirt if you get too many button shirt and wrinkles.

• The stains could simple be better. Need improvement.

• White shirts are a BIG challenge. Use “Black dye, black paint, black smudges” on the negative. It helps a lot.

Some settings: I actually used onetrainer for this version with masked training to learn more of the armpit area. Trained 768 res. 190 images. Epoch 40 from 68. Adamw, 2 steps concept. constant BATCH 4, rank 8/4, no regularization image

Hopefully you can leave some results and some comments. Any idea is appreciated. Thank you.