Description:
I made this using every FFT job illustration by Yoshida Akihiko (missing male bard but I'll include him in a future update). Most of the primary character designs are included in the training data as well.
I made use of the Topaz upscaler software so I was able to remaster many of the images that are otherwise very low res (you lucky b*****ds); even Duke Goltanna...
I made color variations for some of the image flips so the AI doesn't pump out the same colors (blue and gold knight with the 'knight' prompt for example)
Prompt suggestions:
This is trained on anythingV3_fp16 so you'll want to use Danbooru tags in your prompts. Some of the regularly used tags are:
1boy, 1girl, blonde hair, orange hair, light brown hair, braid, ponytail, cowlick, armor, black armor, gold armor, helmet, winged helmet, couter, cuisses, poleyn, greaves, sabaton, gauntets, pauldrons, shoulder armor, hat, bonnet, apron, unitard, leotard, knight, witch, wizard, ninja, samurai, robe, dress, pleated dress, coat, shirt, hakama pants, hooded sweater, tunic, thigh boots, gloves, cape, buckle, ribbon, hair ribbon, sash, loose belt, bow, jewelry, bracelets, scowl, looking to the side, arms at sides, clenched hand, clenched hands, legs apart, holding sword, holding katana, holding cane, bare legs, bare shoulders, cleavage, abs, bald, old man, beardI also manipulated some of the female designs to have exposed breasts so try the following:
breasts, breasts out, nipples, nude, stomachUse the following tags to determine the type of shot (you'll probably have the best luck with upper body):
portrait, profile, upper body, full body
Other details:
Suggested LORA weight is 0.8
The majority of the training data images are 512x768 but some of the portraits were sized to 512x512.
Almost all of the training data uses a white background but I may do a version in the future that features job illustrations with backgrounds.
You get some amazing line art details with hires fix x1.5-x2 but the hands often look wrong at higher denoising strength.
Sorta works with non-Danbooru style models; try deliberate_v2.
I've found that it works really well with UnicornMixes-Mix3 with 640x1152 dimensions.
Usually faces will generate without noses (matching the FFT style) but if you get the tiny dot nose, try adding the 'no nose' tag to the prompt.
Description
50/50 LORA merge of two versions of ffstyle_v1.8: one trained on anythingV3_fp16 and the another trained on animefull-final-pruned-fp16 (NAI).
Both versions were trained on the same training data and parameters.
Use a weight of 1.
FAQ
Comments (5)
the nai trained one definitely works by far the best. Still a bit cooked but maybe because you merged it with the anything one.
One suggestion I have one would be to tag jobs images so if you want to have something resembling a white mage you can use "white mage". Furthermore being able to mix two jobs like that (white mage + dragoon) would really be the peak for this project.
He did a really good job with something similar
but it's pretty much the same approach
I was considering doing one with job specific trigger words (like 'white mage' or 'chemist'). With the current versions you can get part-way there by using the danbooru tags that are specific to the job illustration (e.g., 'long coat' will bring in many calculator characteristics and 'spandex' likes to add the male squire's shirt).
The NAI based LORAs work best when you're experimenting with checkpoints that are different than the one you trained it on. To be honest, I still really like the AnythingV3 LORA's output when generating images with AnythingV3; I think the costumes are really unique.
@guyalOfSfere The NAI based LORAs work best when you're experimenting with checkpoints that are different than the one you trained it on.
Yes my friend that's the point, you're not really supposed to generate using those older raw checkpoints nowadays. You use them solely for the flexibility and database they contain while training. Nai has by far the largest when it comes to anime data and danbooru concepts, Anything is a pruned version of Nai so it still has most of the data but it will also overfit your model more. Similarly to how training on aom3 will turn your backgrounds to simple blurry ones instead of the rich detailed ones you can get when using Nai as base.
Also yes^ the job tags would be great since we as users don't really know how you tagged it or how to get the look of a specific job without guessing and trying dozens of combinations from the tags you did provide. There are no examples or spreadsheet that tells us how to prompt something resembling a dragoon instead of a regular knight etc.
@cyan8 "we as users don't really how you tagged it or how to get the look of a specific job without guessing"
Good point. It'd be better to design the LORA based on how it would be most useful to the average user.
Thanks for the info regarding NAI vs AV3. You've inspired me to train another version of the LORA.