PNSLORA: Male Anatomy for Ideogram4
This is a proof of concept LoKr for male anatomy with Ideogram4.
The dataset was captioned by hand and no artificial images were used for training.
Several concepts were trained for this lora (flaccid/erect + uncircumcised/circumcised + amount of body/pubic hair) but only a few are reliable.
**(uncircumcised flaccid rarely appears when prompted, I may need to set a lower LoKr rank for the next version to retain more concepts).
The model was captioned with a main bbox for the subject and a separate bbox for anatomy.
The anatomy bbox works best with the caption it was trained on:
PNSLORA, (small/average/large) (circumcised/uncircumcised) (erect/flaccid) penis, (smooth/trimmed/hairy) pubic area with testicles.
(ex. PNSLORA, average circumcised erect penis, hairy pubic area with testicles.)
The model is sensitive to bbox size, if you find that anatomy doesn't fully appear or is malformed, making the bbox longer/slightly bigger so there is more space to work with can usually fix it.
Use the model with the regular and unconditional model during generation, regular around 0.80, unconditional at 0.20.
Use dual-cfg with around 5 at most to start with, then to 2 - 3.5 for halfway though (0.500).
This is very much a work in progress, the model is very new and there is lots to learn about what works well for it.
Description
This is a proof of concept LoKr for male anatomy with Ideogram4.
The dataset was captioned by hand and no artificial images were used for training.
Several concepts were trained for this lora (flaccid/erect + uncircumcised/circumcised + amount of body/pubic hair) but only a few are reliable.
**(uncircumcised flaccid rarely appears when prompted, I may need to set a lower LoKr rank for the next version to retain more concepts).
The model was captioned with a main bbox for the subject and a separate bbox for anatomy.
The anatomy bbox works best with the caption it was trained on:
PNSLORA, (small/average/large) (circumcised/uncircumcised) (erect/flaccid) penis, (smooth/trimmed/hairy) pubic area with testicles.
(ex. PNSLORA, average circumcised erect penis, hairy pubic area with testicles.)
The model is sensitive to bbox size, if you find that anatomy doesn't fully appear or is malformed, making the bbox longer/slightly bigger so there is more space to work with can usually fix it.
Use the model with the regular and unconditional model during generation, regular around 0.80, unconditional at 0.20.
Use dual-cfg with around 5 at most to start with, then to 2 - 3.5 for halfway though (0.500).
FAQ
Comments (10)
Please share ai toolkit config and the dataset. Would love to learn this.
+1
I shared the ai-toolkit config in the additional files section. As for the dataset, I used the same camera RAWs I used for the Adonis models, I wanted to keep the dataset images as clear and artifact free as possible so the model didn't learn unwanted styles. I captioned the images (100 total) with the official Ideogram llm prompt guide posted on github (with minor changes to include nsfw details/description). I manually removed "AI"-like language ("realistic", "high-resolution", etc.) and simplified unnecessary prompt language, while captioning the anatomy bbox with a standardized list of descriptions based on the state of said anatomy (like the example in the model description posted). I also manually edited the bbox coordinates using kj's prompt builder node for each caption (a process that I dread doing again) so that it was verified accurate. I'll post an example caption used for the dataset below.
{
"high_level_description": "A full-body photograph of a athletic adult male standing on sandy beach dunes with short dark hair and facial stubble, looking towards the camera in natural daylight.",
"style_description": {
"aesthetics": "natural lighting, sharp focus, earthy tones",
"lighting": "soft overcast sunlight, gentle highlights, subtle shadows",
"photo": "medium telephoto lens, eye-level angle, shallow depth of field",
"medium": "photograph",
"color_palette": ["#F5EFE9", "#D4A370", "#8FA16B", "#E8EDF5", "#2F2A25"]
},
"compositional_deconstruction": {
"background": "Rolling sandy dunes covered in tufts of dry yellow-green beach grass. A line of taller green vegetation is visible further back near the horizon under an overcast sky filled with soft white and grey clouds.",
"elements": [
{
"type": "obj",
"bbox": [761, 370, 900, 521],
"desc": "PNSLORA, average uncircumcised flaccid penis, trimmed pubic area with testicles, down orientation. Front side view."
},
{
"type": "obj",
"bbox": [10, 294, 1000, 874],
"desc": "A athletic adult male standing upright facing slightly right but looking towards the camera. He has short dark hair and facial stubble. His torso is bare showing defined pectoral muscles, abdominal definition, and arm musculature with visible moles on skin.",
"color_palette": ["#F5EFE9", "#D4A370", "#8FA16B", "#2F2A25"]
}
]
}
}
"Base Model" -> "Ideogram"
The Ideogram 4 model category can not be added to nsfw loras
@Degenerator123 hope is a thing from civitai team and not ideogram since we are producing a lot of loras and images nsfw from the base model, and i sincerally want to upload a lora using the base model tag
@flusk It's from Ideogram. NSFW LoRAs are prohibited by the model's license. It even asks people to report them. So far Civit hasn't taken any down from what I've seen, so I guess this rule is a concession to try to placate Ideogram.
@xto Damn they really removed the Open in the sourcing
Looks promising. Thanks for posting.