Search Results
6/20/2025, 3:33:09 AM
>>2975224
would definitely second the noob recommendations, though maybe noob cyberfix instead of base personally, bit better anatomical training at the cost of some style bias, similar to though not as extreme as wai (speaking of wai though, if you really want to speed up your SDXL gens and still have decent quality, give wai cheapfast a try and set to like 16-18 steps with euler, made for farming buzz on civit generator but it's still a pretty solid model for local)
wai cheapfast; https://civitai com/models/1291843
noob cyberfix eps; https://civitai com/models/1001795?modelVersionId=1150241
illustrious can still be very good too it's just more for either the experimental stuff (1.0, 1.1, 2.0, etc), or merges with particular style emphasis for if you're going for something in particular, just imo
i know you're not into getting a new GPU but just since someone mentioned it, 3060 12gb should(tm) be a good "budget" starting point, honestly you don't straight up need 12gb as a minimum, some of the weight can be offloaded, but it definitely slows things down a lot and HDDs/SSD's just aren't really built for that kinda workload. Also always like civit's generator (kinda shit but getting better, even if the sites gotten worse otherwise), or NovelAI (needs your money), comfy might be your best bet afterall for AMD gpu's if you really want the one click set up, forge also has one but i think thats just for nvidia/cuda unfortunately and you'd need to do a full install otherwise, could check out stability matrix instead it might have you covered since it can autoinstall all the major UI's, but not used it. if you're deadset on comfyui, try tiled vae decode instead of regular vae decode node for the img2img stuff, unfortunately it won't let you do it for adetailer/facedetailer (iirc) which is a bit of a problem for comfy+low vram, but i don't think that's intended, might have just been my older version.
best of luck
would definitely second the noob recommendations, though maybe noob cyberfix instead of base personally, bit better anatomical training at the cost of some style bias, similar to though not as extreme as wai (speaking of wai though, if you really want to speed up your SDXL gens and still have decent quality, give wai cheapfast a try and set to like 16-18 steps with euler, made for farming buzz on civit generator but it's still a pretty solid model for local)
wai cheapfast; https://civitai com/models/1291843
noob cyberfix eps; https://civitai com/models/1001795?modelVersionId=1150241
illustrious can still be very good too it's just more for either the experimental stuff (1.0, 1.1, 2.0, etc), or merges with particular style emphasis for if you're going for something in particular, just imo
i know you're not into getting a new GPU but just since someone mentioned it, 3060 12gb should(tm) be a good "budget" starting point, honestly you don't straight up need 12gb as a minimum, some of the weight can be offloaded, but it definitely slows things down a lot and HDDs/SSD's just aren't really built for that kinda workload. Also always like civit's generator (kinda shit but getting better, even if the sites gotten worse otherwise), or NovelAI (needs your money), comfy might be your best bet afterall for AMD gpu's if you really want the one click set up, forge also has one but i think thats just for nvidia/cuda unfortunately and you'd need to do a full install otherwise, could check out stability matrix instead it might have you covered since it can autoinstall all the major UI's, but not used it. if you're deadset on comfyui, try tiled vae decode instead of regular vae decode node for the img2img stuff, unfortunately it won't let you do it for adetailer/facedetailer (iirc) which is a bit of a problem for comfy+low vram, but i don't think that's intended, might have just been my older version.
best of luck
Page 1