And here we have another LoRA from the Delorean series. This time...1980's aerobics! Trashy, rad, ugly pictures of poor quality like cut out from an 80's magazine.See the version description. It describes the changes from versions 1 and 2. The main difference is that LoRA works better with cfg&lt;7 (4 is the optimum).Also... If you like my work, I'd be glad if you would tip it with a small amount to help me partially cover my expenses with Google Colab. Thanks a lot! <a target="_blank" rel="ugc" href="https://ko-fi.com/stepaan">https://ko-fi.com/stepaan</a><h3 id="heading-595">How to:</h3>Lora is trained on period pictures from sports and fashion magazines. I certainly didn't want to enhance them in any way (OK, I did enhance them, but only enough to emphasize the main point). My goal was to capture the style of the photos as much as possible. So don't expect crystal clear pictures! On the contrary, they will be full of noise, scratches and with strange colors.So... :<ul><li>I recommend the cleanest checkpoint possible. Stylized models aren't retro enough ;-). As always, analogMadness is the obvious choice. (From v3.0 it's possible to use the checkpoint of your choice. Still applies, you might get images, that are far from the 80's.)</li><li>Anyone who doesn't like strong stylization and low quality (understand retro) results can reduce the strength to taste (0.5-0.7), but once again, LoRA is supposed to work at full strength (1.0). If you get low quality images, go find a better checkpoint. ;-)</li><li>LoRA knows quite a few types of aerobic clothing. That's why I recommend using Cutoff Regions (I use ComfyUI, I hope that's what it's called in A1111).</li><li>I recommend dpmppsde karras. Or use the sampler that suits your imagination. Also this VAE <a target="_blank" rel="ugc" href="https://civarchive.com/models/88390/difconsistency-raw-vae-pack">https://civarchive.com/models/88390/difconsistency-raw-vae-pack</a> helps to achieve more desaturated retro feel.</li><li>HighResFix is a must, Face Restore is a no no!</li></ul><h3 id="heading-596">Sample prompt:</h3>See images. The Lora can produce huge variety of images (quality wise, retro-ness wise).Enjoy the atmosphere of sweaty gyms from 40 years ago. More Deloreans coming soon ;-)

Delorean-80s-aerobics

Further changes in the dataset (images tweaking, tags changes). Larger network (again, but last time). The LoRA now works with cfg 4 (cfg around 7 is too strong already). Also works with more diverse checkpoints now.

Files with SHA256: 0c93d7fd9a55d1110e0339d67657283c8af63bc688c3480c31c589f85a11174b