v4.6

Added 56 new images to dataset of v4.0. Mostly generated from random prompts from civitai feed. Included different styles (sketch, painting, flat color) to try to fix style overrides.
Changed clip skip to -2 while other params are same as for v4.0.
{
"lr": 0.00015,
"engine": "ai-toolkit",
"epochs": 10,
"ecosystem": "sdxl",
"keepTokens": 0,
"networkDim": 64,
"numRepeats": 2,
"resolution": 1024,
"lrScheduler": "cosine",
"minSnrGamma": 5,
"noiseOffset": 0.15,
"networkAlpha": 64,
"optimizerType": "adafactor",
"shuffleTokens": true,
"textEncoderLr": 0.00005,
"flipAugmentation": true,
"trainTextEncoder": true
}

v4.5
Another experiment. Completly new dataset of 49 unique synthetic pictures. All with different styles and mostly different poses. All with color inpainted.
Results are strange and a bit unstable. This is mostly because I messed up training params and lora got completly overcoocked, showing black squares already on epoch 7. Main file is actually only an epoch 4, because it looks best out of all of them. Main goal was to fix style overrides, but I doubt that I succeded in it.
Whould appreciate feedback on that topic in comments.

v4.0

Another revise of the dataset. Apparently, qwen image edit is usable with my machine and not even trying to blow it up. So now dataset is smaller (95 images, 28 per concept + 7 with none) but with effect of glow more pronounced. Training parameters is as follows: UNet LR=0.0003, Text Encoder LR=0.00003, NetworkDim=64, NetworkAlpha=80.
Colored perineum and spongiosum are gone. Sad but necessary. May be they will go to separate models, but here they are making more noise.
As for new tags - "gt-color" and "gt-bloom" are self explanatory; "gt-dias" is for transillumination but not as good, because textures are weak and they are the main point of it. Transillumination (also nown as diafanoscopy or "hey, my fake breasts are glowing") may also go to a different model later, as it requires higher level of details and much more headache with qwen.

v3.0

More images. Size of dataset is grown up to 155 pictures. For this version Network Dim 32 and Alpha 64, so influence of LoRA is moved towards earlier steps. Considering that model basically generates big splat of color at first and draws details over it later, now this splat is bigger and more vibrant. Tags for colored corpus_spongiosum, perineum and urethra now are also more prominent. Still unstable, but at least it is possible to see them.

v2.0

Generated 9x9 additional images for dataset with only color being different. More importantly - used Network Dim 32 and Alpha 48. Results are interresting. Model now deffinetly better localizes color and has less impact on default style (less tags needed to generate image). For cases, when color leaks on penis, use "pink_penis" tag. Model can not generate pink color for now, so defaults to normal skin color, wich is in some way also pink.

Concepts of colored urethra and perineum are still in there, but haven't been touched, so most likely just got weaker.

v1.0
First try. 95% of dataset is syntetic with mostly automatic tags (not so accurate). Colored testicles do work, but often still colors penis also. Dataset does have some amout of more exotic tags, such as "*COLOR*_glowing_urethra" (technically it is not urethra but "corpus spongiosum", but I don't belive, that normal users do know what that even is) and "*COLOR*_perineum". They are unstable and not with all colors, so more for experiments for now.

Base
I have seen couple of LoRAs for making glowing penis and other parts. But it was always either only for penis, or for whole genitalia (cock and balls both).

So, main goal of this lora is to be able to color specifically testicles, not touching anything else. Dataset mostly syntetic, so no varaity for textures or patterns. Also, untill I get more good images with good-looking glow effect, glwing and just colored testicles will look similar.