Play with weights, the cartoon style may clash with pony's anime bias
It can make some weird mouths
Below is a list of keywords used during the training of this model:
[20830] parody
[15140] solo
[14996] 1boy
[9938] black_hair
[9484] 1girl
[8744] dark_skin
[8720] teeth
[8676] open_mouth
[8545] short_hair
[7661] shirt
[7426] smile
[7186] style_parody
[7019] green_eyes
[6263] indoors
[6109] upper_body
[6021] dark-skinned_male
[5993] blue_eyes
[5737] multiple_boys
[5403] no_humans
[4568] hat
[4476] gloves
[4471] meme
[4452] 2boys
[4146] white_hair
[3982] makeup
[3621] long_hair
[3595] pants
[3481] outdoors
[3324] lipstick
[3295] facial_hair
[3035] bangs
[3033] purple_eyes
[3028] bodysuit
[3020] t-shirt
[2982] holding
[2961] clenched_teeth
[2952] sky
[2808] standing
[2750] looking_at_viewer
[2744] white_shirt
[2656] multiple_girls
[2521] grin
[2423] red_eyes
[2374] blunt_bangs
[2290] closed_eyes
[2284] simple_background
[2264] weapon
[2206] short_sleeves
[2193] cloud
[2170] building
[2165] crossover
[2115] english_text
[2103] grey_hair
[2083] jewelry
[2079] pokemon_\(creature\)
[2026] 2girls
[2001] night
[1961] colored_skin
[1860] orange_hair
[1854] bag
[1821] boots
[1795] frown
[1761] purple_background
[1751] brown_hair
[1739] robot
[1731] dark-skinned_female
[1719] glowing
[1707] skirt
[1705] outline
[1703] spiked_hair
[1699] blonde_hair
[1647] full_body
[1584] closed_mouth
[1577] black_gloves
[1559] long_sleeves
[1544] backpack
[1532] looking_at_another
[1531] midriff
[1496] window
[1488] yellow_shirt
[1485] jacket
[1467] tree
[1395] formal
[1340] suit
[1324] border
[1322] shoes
[1311] bob_cut
[1305] sitting
[1241] green_background
[1237] mustache
[1212] tank_top
[1209] close-up
[1203] blue_hair
[1143] 3boys
[1128] earrings
[1120] purple_lips
[1110] crop_top
[1102] wide-eyed
[1085] necktie
[1060] white_gloves
[1038] red_headwear
[1038] cosplay
[1020] plaid
[1007] jeans
[1006] tiles
Description
FAQ
Comments (8)
Does 20830 mean it occurs 20830 times in data and you used (at least) tens of thousands of screenshots to train this?
Yes. FFMPEG's the entire series at about 1 screenshot every three seconds then let a 4090 run it overnight
I'm aware its overkill
@Synthdark8 Wow. And here I thought the 1000 screenshots with hand tagging (wanted to recognize some characters/concepts as well as style) I was working on would be a lot.
@NanashiAnon 1000 Handatagging is still quality work. I would even question if this insane overkill amount of images that OP used, did really make a significant change in quality.
@vynenave926 If nothing else, it did make it able to reproduce a lot of characters. No auto-tagged Gundam X style to compare it to.
@NanashiAnon Not questioning that it worked in the end. But whenever I trained styles, sometimes even 60 images were enough, it had to be 60 high quality images, but they were enough. I did handtag them though, so error in tagging was not a problem. ; But the real question here, would be if OP used only half of the images, or even less, would there be a significant change in the result? If for a good amount of styles less than 100 images work, is the amount OP used completely contributing to the final result or practically just a placebo?
Like when SD started to become more common, someone created this extremely long negative prompt, that was said to create perfect hands and a lot of people believed it, because it theoretically got the right result, but it wasn't because the negative prompt was a miracle. The one who created the prompt didn't understand how prompts really worked and was adding so many tags until it finally worked. Not because all the tags were taken into account, but because some of the tags were directly working against each other and the token limit only used the stuff that would blend together.
@vynenave926 I originally started by doing 30k images/steps but I've cut it down to 20/15K without a significant drop in quality. 10k has worked too (See Lolirock) but I still like generating screenshots off the entire series and I'm not tagging 10-30k screenshots by hand.
@Synthdark8 Thank you for the information :) , I tested this lora today and it works really well.
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.



















