ScytheXL

DISCLAIMER: This is my first attempt at training for Stable Diffusion. Despite all the time and effort put into it, I consider this version of ScytheXL, v0.24.224.05, a prototype. So you may encounter some odd results and/or other bugs during use, ones I will do everything I can to resolve for future versions of ScytheXL. That said, thank you in advance for all for your patience, as well as for any help/improvements you may be able to provide!

TIPS: Ko-Fi

QUICK DETAILS:

• ScytheXL is a simple SDXL LoRA made to generate scythes and characters holding scythes using Stable Diffusion's SDXL 1.0 models.

• This LoRA was trained on SDXL 1.0 Base, and then two versions with different desired strengths/weaknesses were merged into each other to stabilize the resulting output.

• As such, this LoRA seems to work on most realistic models of base SDXL 1.0 heritage I've tried. But unrealistic SDXL 1.0 models produce some results that are... interesting.

• I trained this LoRA on the base SDXL VAE, so I've been using that during testing. Although I prefer Sharp Spectrum and Flatpiece VAEs over the base SDXL VAE.

• It is imperative that you use the trigger word or phrase if you want this LoRA to work, these being "scyth3" and "holding a scyth3". Very important!

• The strength settings I used varied by model but, typically, sat between 0.65 to 0.85. Usually 0.7 or 0.75 works perfectly. But I varied this more for unrealistic models.

• I went through pretty much every sampler available in Forge during testing. This LoRA performed reasonably well all around. So my sampler recommendation is "dealer's choice".

• CFG isn't something I normally play with, so I've just been going between CFG 6 and CFG 7. Another "dealer's choice", if that's something you use.

• Hires Fix and ADetailer were used with sample images. I used Foolhardy's Remacri at 2x with Denoise at 0.39 for HRFix, and then ADetailer on face and hands.

THE STORY:

This project started in June 2024 while working on another Stable Diffusion related project involving a character that was meant to wield a scythe. A plain, traditional, reaper-esque scythe. I started with some GPT-infused, next level prompt engineering, and then I tried looking for LoRA models and Embeddings meant to generate scythes, but none really existed. Eventually I came to the realization this just couldn't be done with the tools available. At least, not yet.

With a group of 20 awful photos pulled from Google, I started training LoRAs with pretty terrible results, but just okay enough to develop a viable dataset with. After a lot of (re)training and grinding, I ended up with 100 unique, high quality images in total. With that new dataset, I started training variations of this ScytheXL LoRA, in an effort to find which combination produced the best LoRA, until I finally settled on the LoRA I've uploaded here.

However, this LoRA model is more of a proof of concept or prototype than a beta, one that is very much in its infancy. Therefore, YMMV. But I do want to thank everyone, in advance, for all the support and patience they afford me and my model!

I wish you all the very best and lots of AI fun!

- Dom