Please download at least one of my images and drop it into ComfyUI to see my workflow
Trained using my own 3D Rendered images with (mostly) my own devices, some of which I share on my DeviantART account https://www.deviantart.com/squealnsquirm/gallery/71668896/3d-assets-for-daz-studio
Training involves having “, 3D Render” at the end of every training prompt for 3d rendered images, then simply removing “, 3D Render” will result in the model+lora producing a photographic looking image instead, the idea being that the model will identify “3D Render” as a style.
This mostly works, but isn’t perfect, composition (sometimes) requires “, 3D Render” to be present, so I developed a split generation workflow in which the first half can concatenate “, 3D Render” with the positive prompt, but the second half will not use this tag, so you can have the 3d rendered composition without the 3d rendered style.
I also tried training using generated images to remove “, 3D Render” entirely, but wow did that that fail badly, turns out models do not like being trained on their own generated images.
You may notice some of the images provided are anime style, ahh… well, I never trained it to do this, just part of the incredible robustness of ZImage Base. It’s Amazing!
Device, Concepts and Formatting
This is the formatting used to train, thanks to ZImage Base model, it is possible to break this formatting and still produce good results, stuff I never trained it to do and it still produces predictable good results, honestly amazing, beyond what I would have though possible. My earlier fine tune of SDXL used the same formatting, but if you broke the formatting at all, it wouldn’t work. ZImage is a lot better.
If you are confused, please download some of my examples
naked person in material right hand over left armbinders with material bolts, hands, location, angle
naked person in material reverse prayer armbinders, hands, location, angle
naked person in a material bodybinder, hands, location, angle
naked person in a material upright pillory, hands, location, angle
naked person in a material mobile pillory, hands, location, angle
naked person bent over in a pillory, bent over pillory with material frame material padding and material head board, hands, location, angle
naked person with {an erect|a flacid} penis
additionally
material ponygirl belt with __hair_color__ tail and materialbuttplug
material ballgag
material leg irons with material chains
fields
person = {girl, man, woman}
material = e.g. metal, pink plastic, cyan metallic, military camouflage, transparent… many more
hands = {hands curled up into a fist, hands open fingers curled, hands open fingers curled}
location = where the subject is e.g. outside a shop in the city
__hair_color__ = blond, brunette, light brown, colorful
[1] hand over hand armbinders can be formatted in many different ways
[2] bent over pillory is fiddly, you need to match the formatting and you need the first half of the rendering to use “, 3D Render” tag, am still working on getting this working. Basically didn’t work at all with SDXL, so still a huge improvement.
angle
3D Renders were created using a rotating camera rig in Daz Studio, there are 12 angles in total
these angles can use multiple angles in a single line with the most significant being at first
e.g. “from the front and from the left” will be more from the front then from the left
may also have “from above” or “from below” and may be terminated with “, full body” for full body view
left and right is relative to the person not the viewer, e.g. “naked girl, from the left and from the front” would have her left hand side in the foreground
from the front
from the front and from the left
from the left and from the front
from the left
from the left and from behind
from behind and from the left
from behind
from behind and from the right
from the right and from behind
from the right
from the right and from the front
from the front and from the right