>>5178

...There is only 1 good branding iron on sketchfab though. But if you want to make custom branding marks you'll have to make the stamp part yourself anyways. Then all you need is a rod with a wooden grip and a hand.

>...pears of anguish, spike covered dildos, needles, barbed wire, scalpels, saws etc...

There are 2 entries for pears of anguish on sketchfab. Spike covered dildos will necessitate a bit of work again, I'd think, but at least there are a handfull of dildos there, so that's a start. If you just type in torture into the search you'll find plenty of interesting stuff there.

>here's the original template image btw, from homerus.

lol, I've never seen that pic, but by sheer coincidence choose the same type of rifle as the original artist. That's definitely an AK, too.

>haven't done much inpainting yet (need to spend some time finding the right settings)

In a1111 sd-webui/forge what has worked best so far for me is: Leave Resize mode as "Just resize", set Mask blur to about double the defaults, so 8-10, set inpaint area to "Only masked", Only masked padding pixels again to at least double the default value. (I'm not sure I'm understanding it 100% correcctly but from what I think I know this value determins, how far outside of your mask the AI will consider the rest of the image. to blend in/make a sensible inpaint. So sometimes blowing that value up to the extreme can help to force the AI to stop messing up, but in most cases something between 68 - 92 pixels should be sufficient.) Next check "soft inpainting" and leave all values therein as they are. Don't forget to check your sampling method (when sending a pic from txt2img to img2img that setting doesn't get transferred). When you mask the object/part of the image you want to change you should draw the masking area with a high overlap. For instance when I inpainted just the straight part of the branding rod on this pic here the mask was about 3-4 times as thick as the rod itself. When you're done with masking you'll have to eyeball (guess) the width and heigth of your masked area (and add some extra space, just to be sure.) Then in the "resize to" column you can now choose these estimated values. This way, instead of loading/denoising the whole picture, only the chunk of the image containing your mask will be loaded and worked upon. Its much much faster and efficient this way. Remember that everytime you send an inpainted pic from the right side of the ui to the inpaint window the "resize to" value gets reset to the images values. Also don't forget to check "inpaint area: Only masked", otherwise these setting will rescale the output of the whole image. I'd recommend to not go under 596x596 pixels for SDXL based models.

The CFG scale and Denoising strength you will most likely have to experiment on a lot at first. For starters try the cfg at something like 3.5-5. Depending on the size and type of inpaint subject you may sometimes need only very little denoise, for instance if you're just fixxing some eyes something under 0.4 should be sufficient). You can (and will need to) go higher in many cases, but without controlnet with above 0.55 denoise it will start to generate more and more nonsensical stuff. With the aforementioned setting (and of course depending on your hardware) doing an inpaint run for 10 or so outputs will only take minutes. So experimenting shouldn't hurt too much I hope and sooner rather than later you'll get sort of an intuition for at least the right ballpark area for these settings.

Promptwise you can leave the "score_x, source_bla"- part in and then just describe what the inpainted area is supposed to become. If it is a part of a human (/creature) - subject prefacing the prompt with just the word "perfect" has helped me greatly. (i.e. perfect blue eyes of girl, ...)