hello y’all ✨ i finally finished a proper version of the model that i have been working on for the last weeks. this was a bit more ambitious than my previous models. it was finetuned on a few hundred & mostly hand-captioned images at higher resolutions than the usual 512×512 of the SD 1.5 base. the vibe is painterly & ethereal i guess?
i released an earlier version here.
examples / comparisons
one-shot and unedited/not post-processed
as you can see from the example images, this model does not require trigger words.
if the eyes are too borked for your taste, it can help to raise clip skip to 2 (default is 1). you find it under Settings → Stable Diffusion in the A1111 web ui. i recommend just adding it to the main tab because it’s super handy. to do that, just add “CLIP_stop_at_last_layers” to Quicksettings list in Settings → User Interface
here’s a screenshot of the settings I usually go with:
- it should include the newest VAE (vae-ft-mse-840000-ema-pruned.ckpt), but if your results look a lot worse than the example images above you can try turning it on manually in the web ui settings. this explains how to
- all examples are made with “Use old karras scheduler sigmas (0.1 to 10).” activated under Settings → Compatibility
- i recommend adding “nude, naked” to your negative prompt if you don’t like boobas because this model certainly does (￢‿￢ )
update: these also seem to work well with this model. weapons do look borked tho
images that look more like vector art can be prompted with “illustration”: