This LoRA allows you to create civilian persons wearing cloths as it was common in ancient Rome.

The training had images that were captioned with these trigger words:

• tunica laticlavia - a tunic with two wide stripes (clavi)

• tunica angusticlavia - a tunic with two small stripes (clavi)

• tunica recta - a white tunic without stripes

• tunica - any other tunic without stripes and in any color

• exomis - similar to a tunic, but with one shoulder and the chest exposed, usually worn by man doing hard work

• toga praetexta - a toga with a wide border line

• toga trabea - a purple toga, possibly with scarlet stripes or elements, worn by augurs

• toga pura - a toga without a border line

• toga exigua - an earlier, less voluminous style of plain off-white toga, shorter than the later imperial toga pura, often worn without a tunic and draped to leave one shoulder and part of the chest exposed

• stola - a long, often sleeveless, ankle-length dress worn by Roman matrons (woman), typically over an under-tunic

• palla - a large rectangular shawl of various colors, draped over a stola (or tunic) by women

• braccae - trousers of various colors and lengths, unusual

• calcei equitati - red, black or red and black boots that were worn by senators

• calcei patricii - brown boots

• carbatina - simple, common shoe

• soleae - sandals

• sagum - a rectangular woolen cloak

• paenula - a heavy, often conical or bell-shaped cloak, closed or fastened at the front, typically but not necessarily with an attached hood

• subligaculum - a loincloth or simple undergarment worn by both men and women

• strophium - a band of cloth or soft leather wrapped around the breasts for support, worn by women, particularly during exercise

Prompting notes:

The training did pick up the differences, bit they are sometimes weak. So experiment with the trigger words and sometimes it might help to add actually redundant information like prompting a tunica laticlavia and also prompt the stripes or clavi.
The Krea 2 version has much better prompt adherence and output quality, i.e. usually it is sufficient to prompt normal.

The training was done with images in the following formats with a focus at 1:1 and also many 2:3 and 3:2. To get best results you might want to use these dimensions as well.

• 768x1344 - 9:16

• 832x1216 - 2:3

• 896x1152 - 3:4

• 1024x1024 - 1:1

• 1152x896 - 4:3

• 1216x832 - 3:2

Training details

All training was conducted with SimpleTuner.

Krea 2:

• When you want to use it with ComfyUI, please use https://github.com/Comfy-Org/ComfyUI/pull/16857 (and give that PR a 👍 so that it gets pulled quickly);
  using it with diffusers should work out of the box.

• LoKR with factor=12 (Krea2Attention=12, Krea2SwiGLU=12), linear_dim=10000, linear_alpha=1, use_scalar and full_matrix as well as DoRA (dora_wd=true)

• 178064 steps (22258 real steps * batch size 2 and gradient accumulation 4)

• Trained with Krea 2 Raw, to be used with Krea 2 Turbo

• training time: 55.3 hours on one 5090, including validation image generation

• regularization images were used

• masked training with the faces masked away

• each image had multiple captions for versatility