Search Results
6/29/2025, 3:17:23 AM
>>105739030
yeah the anatomy is a pain point but it's a big step up from 3.0 Medium IMO. That said sampling with Skip Layer Guidance and Clownshark "res" samplers (I like res_3s) helps a ton in my experience.
Also it's not a finetue but I've been iterating on a big SD 3.5 Medium detailer Dora for a while that I'm ALMOST happy with at this point. It doesn't completely fix everything people wise (partially becaue it's not really specifically realism-focused, it's more of a mixed dataset that also includes a lot of 2D / 3D content), but can at least make "same seed / same sampling settings / same prompt" improvements like here (stock medium on Left, with Dora on right, and yes I know her right thumbnail is a bit wack still) as of my latest testing version of it.
Everything I've learned about training 3.5 Med in the process is useful though for if I wanted to do a separate one later that actually was only trained on photos and nothing else (which I might)
yeah the anatomy is a pain point but it's a big step up from 3.0 Medium IMO. That said sampling with Skip Layer Guidance and Clownshark "res" samplers (I like res_3s) helps a ton in my experience.
Also it's not a finetue but I've been iterating on a big SD 3.5 Medium detailer Dora for a while that I'm ALMOST happy with at this point. It doesn't completely fix everything people wise (partially becaue it's not really specifically realism-focused, it's more of a mixed dataset that also includes a lot of 2D / 3D content), but can at least make "same seed / same sampling settings / same prompt" improvements like here (stock medium on Left, with Dora on right, and yes I know her right thumbnail is a bit wack still) as of my latest testing version of it.
Everything I've learned about training 3.5 Med in the process is useful though for if I wanted to do a separate one later that actually was only trained on photos and nothing else (which I might)
Page 1