>>2129

>I'll test that lora, haven't tried it yet but seems neat.

Oh, my mistake then. I thought I had seen one pussystabbing by you in the other thread but that must have been someone else's work then.

>>2130

Oh, but yes, there already is a Lora! >>1800 That was trained by Entropy with the dataset I made. Or to be more precise, I made/collected the pics and then gave it to one anon, who then gave it to Entropy, who then captioned and trained it with his hardware.

So far knives work best, rapiers not at all and swords kind of. (Blades are present, but of course SD mostly messes up the hilts) But I believe that most of that can be easily fixed. There are medieval swords with classics crossguards as well as ancient roman gladius style swords with that stubby peppermill style hilt in the dataset. Since both got captioned as "sword" the ai doesn't recognize the difference. Same happened with rapiers and sabres. But to be clear that's my fault. When I send that dataset to said anon I had captioned only a tiny portion of the images... poorly... I should have organised the pics according to their type to differentiate. Also somehow around 40 pics went missing somewhere along the way. I suspect I made a mistake when zipping the archive. So a V2 of the Lora will most likely be even better than it already is.

>Although I wonder if odd unique swords such as rapiers wouldn't contaminate the generation process, being that guard construction is always so unique, I question whether it can be easily averaged.

Yeah I suspect that, too. But since they just fit perfectly for the kind of scenarios I'm most into I just have to give it a try and see if it can't be done somehow. I'd even settle for simple cuphilts. But in any case there are far too few examples of real rapiers in the dataset as of now. But at least sabres I believe should become possible with a retrained version.

In general stylebleed is also a bit present with the Lora, because more than half of the pics where made with that western artstyle that is also most present in my "pussystabbing" thread. So having more examples but with other artstyles should also greatly improve the Lora. :)

>Thanks, I saw your last gens, very clean and nice with the upscaling. Multi diffusers?

Oh, now I have to out myself as a noob- hack- fraud again... what are multi diffusers? :o ...Basicly my workflow concerning upscaling is as follows: find the gen I want to refine, run that prompt and seed with Hiresfix upscale by 1.25times @0.24 denoise, 16steps, either ScuNet Pnsr or RErsgan anime. Then I do all the inpaint shenannigans and finally upscaling in the extra tab with 4xUltrasharp as 2nd upscaler, 2nd upscaler visibility at 0.3, GFPGAN at 0.3 and then I'll do 4 passes, one each with ScunetPSNR, RErsgan Anime, RErsganx4 and lastly 4xUltrasharp as 1st upscaler. In almost all cases I could just do 4xUltrasharp and 0.3 GFPGAN, most pics I upload here are done with that. But in some rare cases I like the slightly washed out look of ScunetPSNR over the others or one of these upscalers may mess up some detail. So I have 4 upscaled versions to choose from. Bit inefficient but it just stuck with me. xD