Are you tired of SD giving you this completely prude tight fit high-cut armholes tank tops? Do you want your generated dork himbos and bimbos showing off their muscles at the gym?
Let’s fix it!
Lora weight works from 0.3 up to 1.0.
Start with 0.68
There is no trigger word. But you should reference a tank top, sleeveless shirt, gym stringer or singlet. If for males you are getting too much belly crop top shirts or shirtless it helps to use “Shirt:1.2” and just “tank” without the word “top”.
Supporting prompts:
shirt, tank top, tank, sleeveless shirt, (high crewneck:1.3), low-cut ????? (males), low-cut cleavage (female), extreme low-cut dropped side holes, side ????? (female), side pecs (males), (male/female) ?????? slip, side view, profile, front view, full body medium shot, long shirt, loose shirt, muscle body, thin body
Negative supporting prompts:
skin holes, shirtless, jewelry, necklace, muscular, head out of frame
I did a lot of testing without the lora and SD just don’t know the concept of a shirt “armhole”. So, there is no point in using this word. “Dropped side holes” gave the best results comparing to a lot of other combinations I tried. Still, it can’t really do large armholes without this lora.
Also, the unfortunate side effect of “side holes” is holes appearing on other places, like the shirt, or in the worst scenario, the skin (nightmare). That is why I suggest if you really want a “extreme low-cut dropped side holes” (don’t use the word armhole) to use “skin holes” on negative.
SD does not know what a shirt “neckline” is either. So, there is also no point in using that word. To get a high-cut neckline, the best combination of words in my testing was simple “high crewneck” with 1.3 weight. The normal SD will understand that one without the lora. But with the lora it’s way easier. You might get the effect of those manual ripped cropped shirts that get cut to be a tank top. You might also get a sleeveless turtleneck shirt.
In regards to the concept of large armholes itself, this lora is kind of cheating. It is a specific type of clothing lora, not a concept. Maybe 2 clothes (low-cut cleavage tank tops and high-cut neckline tank shirts). It’s not a lora on the concept of “armholes”.
I got quite curious about how to train a concept, not a clothe like this one. I might try another lora that have a different dataset on the concept of “armhole”, so maybe we can adjust any type of clothes armhole sizes with that other lora. But maybe. It will probably be hard to train this kind of abstract specific concept.
Even though I hate training Textual Inversion Embeddings I thought that because this was a simple concept, it would be nice to try again. It was OKish and as always took a lot of work. It can be used to support this lora if used together. But honestly, the lora works great on its own. I did not test it but maybe the TI+Lora together might make it more flexible. I’ve seen it happen for my “Fares Fares” lora+TI.
I did a SDXL version already that works quite well.
_________________________________________________________________________________________________
Try out SD15 and TI here (will update later):
Stringer Gym Tank Top / Dropped Side Holes / Large Armholes – SD 1.5 Textual Inversion
Stringer Gym Tank Top / Dropped Side Holes / Large Armholes – SDXL
_________________________________________________________________________________________________
The base checkpoint is the original “v1-5-pruned.safetensors” 7.7GB. So, it should be very flexible with any checkpoint. I recommend realisticVisionV51 (also the inpainting version), Reliberate_v20, objectiveReality_v20
CFG: 5.5 - 9
Clip skip: 2 sometimes works more consistently with one color, but 1 is better for general realistic pictures.
DPM++ 3M Exponential (50 steps or more)
DPM++ 2M Karras (25 Steps or more)
DPM++ SDE Karras
DPM++ 2S a Karras
I could not find a good lora block weights yet, but I didn’t do a extensive research on it. it might improve results.
Because the base and the contours of the shirt and body is created in the beginning og the image denoise, Loractl is a GREAT option if you want the lora to not interfere too much with facial structures and details that get added latter for a specific character or celebrity Lora. I would recommend to star high weighted (0.8) at 0 steps. At 40% of the steps drop the weight to 0.35 ou even 0.
For example stringer:0.8@0,[email protected],[email protected]
Want to have some “fun”? Install wildcards dynamic prompts extension: https://github.com/adieyal/sd-dynamic-prompts . And wildcard manager. Here it is a good crafted script I made for testing. Paste on prompt:
{side view|from below|[0.2::(front:0.6)]} (medium shot:1.1) of a YetAnotherWildcardCollection-main/States/Nationality {man|woman} {18|28|45|80} years old on a YetAnotherWildcardCollection-main/Background/*, wearing a {dark|ligh| } ({white|black|red|pink|purple|green}:1.1) (tank:1.1) {(high crewneck:1.2)|(extreme low-cut neckline:1.2)}, {(extreme low-cut dropped side holes:1.1)|low-cut dropped side holes| }, stringer (shirt:1.2), { |(?????? slip:1.1)}, natural lighting, 4k uhd, dslr, soft lighting, high quality, Fujifilm XT3
Problems with the current Lora:
- Using “extreme low-cut dropped side holes” will give holes to other things. Use “skin holes” on negative
- Bias to muscular body type. Just prompt it away
- ??????? showing weirdly below the shirt. I always see this with SD, don’t think it’s exclusively a problem with this lora
Some settings: Trained 768 res. More than 205 pictures. 2 steps each image. 3 different folders/concepts (2_stringer shirt / 2_low-cut ????? stringer shirt / 2_low-cut cleavage stringer shirt). AdamW.
Hopefully you can leave some results and some comments. Any idea is appreciated. Thank you.
