Susan, my favorite FJ model, trained at 768*768. If this comes out a bit distorted at 512*512, I can train the same images at 512*512, just let me know? So 20 images, 1 Epoch, 2600 steps. This Lora seems to play really nice with every model I've tried it with so far.
Description
FAQ
Comments (6)
Unfortunately the LORA does not play well with SD Upscale in Img2Img
ok, I haven't used SD upscale, looks like I might have a bit of a learning curve to understand why .You definetly started at 768 x 768? This Lora gives me strange results at anything less
@Hbait2 That may have had something to do with it; really hoping for a future update! You have great taste.
@gopman766 thanks, I'll re-train at 512x512 and increase the number of training image and post another model when done. 768x768 training does seem to give better results
@Hbait2 In my experience, some LORAs can be a little funky looking when first generating at 512x512, but the results come out better when taking the initial output and using SD Upscale with the LORA still enabled.
@gizman667 have run training at 512, the Lora is exactly one KB larger than the 768 version, so Civitai won't allow me to upload here :( - any suggestions?