Stable Diffusion

4309 readers

4 users here now

Discuss matters related to our favourite AI Art generation technology

Also see

Other communities

founded 1 year ago

MODERATORS

db0@lemmy.dbzer0.com

Training Loras - Is SDXL still the best? (poptalk.scrubbles.tech)

submitted 6 months ago by scrubbles@poptalk.scrubbles.tech to c/stable_diffusion@lemmy.dbzer0.com

5 comments fedilink hide all child comments

Hey folks, I've been training Loras now for a while, and have some scripts I really like that I've been working with. However, I realized I haven't been keeping up lately, so, is SDXL still the best for Loras? And by that I mean before with 1.5 and standard SDXL is the most accurate quality I've received.

My Loras seem to work fine with Lightning and Turbo models, but is there anything else I should look into? Any major things that you've changed in the last 6 months to your trainings? Is sdxl_base the best basemodel to train off of?

you are viewing a single comment's thread
view the rest of the comments

[–] wewbull@feddit.uk 7 points 6 months ago (2 children)

Comparing base model to base model I think Cascade is quite a lot better than SDXL, but .....and it's an enormous but... It seems to have been shunned by the community.

Maybe nobody with resources to do training is interested in a model with commercial restrictions, or the multi-model flow was just too different for people. Not sure, but the output of the base model can be really nice. Not always, but I find the biggest errors are people taking on a painterly/waxy appearance rather than the arm turning into a leg body horror you can get with SDXL. I think the "compressed" Stage C works to keep the composition together across the whole image more.

[–] scrubbles@poptalk.scrubbles.tech 1 points 6 months ago

Thanks! Ill have to try it out to see what it takes to train, but their Numbers do look nice

[–] istanbullu@lemmy.ml 0 points 5 months ago

I think Cascade doesn't have the ecosystem that SD15 and SDXL have built in the last year. It might be a better model on fundamentals, but that doesn't help if the ecosystem is not there.