Make sure to use trigger words, "leg stumps" & "arm stumps" as appropriate. In supported versions, "shoulder stumps" and "thigh stumps" can be used for shorter stumps cut off at the shoulder/thigh. "amputee" and "quadruple amputee" were NOT used often in captioning training images, so they may not be very effective.
Recommend using 0.8a if using civitai generator. For later versions, try using locally, or try using a merged model like autismmix. dora versions after 0.8a are mostly just fiddling with certain details: in particular, trying to get individual arm stumps and leg stumps to generate independently and consistently for non-quadruple-amputees, improving stump cover generation, improving chain details, etc. If you just want quadruple amputees, use 0.8a.
Prompting guide:
stumps:
For most versions, "arm stumps and leg stumps" will work best, as the training set has the most examples of quadruple amputees. Other stump combos work fairly unreliably, and may need additional strength (e.g. (arm stumps)).
stump options:
arm stumps and leg stumps
quadruple amputee
works the most reliably
arm/shoulder stumps
amputated arms or shoulders
leg/thigh stumps
amputated legs or thighs
right/left arm stump
amputated arm
right/left leg stump
amputated leg
stump covers (non-exhaustive list):
Stump covers can be optionally specified. Some work better than others, depending on the version.
stump cover options:
bandaged
[chained loops/handles/eye bolts on] flat/rounded [red/orange/yellow/etc.] [metal/cloth] caps on
scarred
eye bolts in
For prompting, combine stump cover options with stump options. E.g. "bandaged arm stumps and leg stumps".
Largely trained on nsfw images, so include specific clothing items in your prompt if you want clothed images. Clothing covering/obscuring stumps will not work as well as sleeveless short-sleeved clothing. Working on expanding dataset for greater flexibility of which body parts are amputated.
Description
slight improvements. Works better locally than on civitai. If using on civitai, use autismmix or some other pony merge to improve image quality (with caveat that lora may work a little less well).
FAQ
Comments (6)
Any chance of you making a LoRA - or whatever's necessary to improve compatibility? This currently works fine with Pony Realism, but with all other checkpoints I've tried, it produces very noisy, disfigured images. This is the only LoRA/DoRA that I've seen do this so it's a bit strange?
I think the issue is that it's overtrained. Though it works much better locally for some reason. I'm working on fiddling with training parameters to hopefully help deal with the issues.
In the meantime, I'd advise trying 0.8a if you're using the civitai generator. If you do want to use 0.9, try it locally, or try using autismmix. tbh, for the simplest case of "arm stumps and leg stumps" images, 0.8a is probably best. All the versions after 0.8a are mostly just fiddling with certain details, in particular, trying to get individual arm stumps and leg stumps to generate independently and consistently for non-quadruple-amputees, getting chain details right, etc.
@out_of_username_ideas this is locally with ComfyUI, for what it's worth.
@nightmare_lead strange. Is it at least better quality than civitai's generator? I'm getting much better results on automatic1111 than on civitai. The gallery images are all generated on automatic1111 with txt2image + hiresfix, bar maybe one or two exceptions that were created on civitai.
What version are you using?
What tool did you use to train this DoRA? I want to make an extremely customizable amputee DoRA for NoobAI-XL like yours for PonyXL. /genq /nm /nf /nav
(no, I'm not requesting a NoobAI-XL DoRA from you, I will do it myself)
I used kohya_ss and fairly standard prodigy training params with an effective batch size of 36 (batch size * gradient accumulation). The customizability largely just comes from the relatively large training dataset size and long-ish training times. My current dataset size is around 800 images or so, and my latest uploaded dora version (v0.9) was trained for 73 epochs. (which also explains why my doras tend to have strong artifacting when used on civitai, which I think does badly with more overtrained loras compared to local software). I also manually tagged with pseudo natural language captioning, with sentences using danbooru tags when possible. (tbh, you'd probably be better off using something closer to regular danbooru tagging).
If you want my training config json, let me know and I'll see if I can find a place to upload it.
I might get around to making a noobai version of my doras eventually btw. A bit wary of doing so now since I'm not sure which big new model is going to be the one with the more thriving lora ecosystem.