This model is a DORA for SDXL. It was trained on the sdxxxl_v30 checkpoint.

The model was trained with 44 pictures found online. The pictures were cropped to remove non-subject matter.

The captions for training were of the form,

"text "THE TEXT" on BODYPART"

Additional captions were added at the end using auto-tagging and manual captioning.

Examples: "1girl,pussy, spreading legs,on bed,breasts,pubic hair"

Bodyparts in caption data are: "upper chest","breast","stomach","pubic mound","leg","ass","navel","upper abdomen","nipple"

Recommended Settings

The model works well with a,

• 0 .9 strength

• 4 CFG

• 30 steps

• DPM++ 2S a Karras

• base prompt "the main prompt for the scene,BREAK text ("THE TEXT":1.3) on her BODYPART <lora:bdwt_sdxl_dora_v1:.9>"

• neg prompt "head out of frame,misspelled words, garbled text" (optional)

Other setting may work as well these gave reasonable results.

Strengths

• Generates bodywriting text

• Controls subject

• Reasonable success rate for correct text generation (1-30%)

• Partially case-sensitive

Weaknesses

• Adds extra text fragments

• Struggles with longer text

• Avoids faces

• Text color is influenced by other colors in prompt

Other

• The images are all cherry picked to different degrees. Do NOT expect 1-shot generation

• The success rate of generations depends on which text is being generate

• Longer phrases have a lower success rate

• Words with repeat letters or the letters "l" and "i" are difficult

• This does not work perfectly, but functions well enough to be useful

• Captions also included "around" or "across" for placement.

Version 2

There is likely to be another version. Additional training data with less words per subject might yield better results. Alternative captioning phrases might work better. If improvements can be made a new version will be released.