Sign In

Bulma Brief (Multiple older versions) - Dragon Ball Z and Super

430

3k

2.5k

57

Verified:

SafeTensor

Type

LoRA

Stats

1,997

2.5k

30

Reviews

Published

Jul 16, 2023

Base Model

SD 1.5

Training

Epochs: 8

Usage Tips

Clip Skip: 2

Trigger Words

buruma_brief_older
buruma_dbz_buu
buruma_dbz_namek
buruma_dbz_cell
buruma_dbz_futuretimeline
buruma_super_start
buruma_super_resurrectionfreezer
buruma_dbz_start
buruma_dbz_garlic
buruma_dbz_afro

Training Images

Download

Hash

AutoV2
D43445E040
default creator card background decoration
knxo's Avatar

knxo

This LORA objective is to produce as many Bulma variations as possible from Z and Super. I was unable to tag for outfits due to how stupidly large it was already. Bulma is just extremely prolific appearance wise having multiple outfits and hairstyles at different points.

I had forgotten the horror of making the first version of this LORA. I guess that's how women feel when giving birth to their second kid. Anyway. The LOCON Version is already out so try both and compare if you want.

V2:

Works fine with:<lora:buruma_brief_olderV2:0.7>,buruma_brief_older, <Bulma version>

Like this: <lora:buruma_brief_olderV2:0.7>, buruma_brief_older, buruma_dbz_buu, standing, solo, large_breasts

The available Bulma versions in this LOCON are:

  • buruma_dbz_start: Weird 80s punkish hair (Works better in the LOCON version)

  • buruma_dbz_namek: necklenght with her hairband

  • buruma_dbz_garlic: longer asymmetric hair

  • buruma_dbz_cell: A little longer than Namek and without the hairband

  • buruma_dbz_futuretimeline: Long hair

  • buruma_dbz_afro: Afro

  • buruma_dbz_buu: bobcut hair

  • buruma_super_start: Bobcut and looks older

  • buruma_super_resurrectionfreezer: similar to start but not quite

I leave to you her clothing tagging. You should be able to get the tags summary inside the dataset so experiment with that.

V1:

Seems people wanted an older Bulma as it seems current LORAS are from her pre DBZ self.

You can see the different variants included in the XYZ graph.

I had ok-ish results in the .45 to .65 range. If you go too low you might be caught in the transition between the normal eye and Toriyama style eyes.

I had best results with: buruma_brief_older, <insert one of the other Bulmas triggers here>, large_breasts, blue_eyes, <lora:Buruma_Brief_older:.5>

Like this: buruma_brief_older, Buruma_Super_ResurrectionFreezer, large_breasts, blue_eyes, <lora:Buruma_Brief_older:.5>

The supported variations of bulma are:

  • Buruma_DBZ_Start

  • Buruma_DBZ_Namek

  • Buruma_DBZ_Buu

  • Buruma_DBZ_Cell

  • Buruma_DBZ_FutureTimeline

  • Buruma_Super_start

  • Buruma_Super_ResurrectionFreezer

  • don't use any of them for a mixed style

I added the tag summary inside the training data.

Creating this was annoying as hell, it came out overfit no matter what i did until i noticed that the noise(Saw a big honking watermark) wasn't from my training data. I cleaned it properly damn it!

Turns out the LORA was getting contaminated with the model's Bulma concept which is fugly and overtrained. This caused my LORA to be overfit from epoch 1.

Long story short after cleaning the dataset and tags and cooking it with different learning rates and coming up empty I found the watermark in the noise.

Finally "fixed" the issue by changing the triggers to the romanji spelling thus why they use Buruma. Also during one of the dataset shiftings ended up tagging everything by saga so you can more or less reliably make the LORA spit individual Bulmas from different story arcs.

Not my best work but one of the more annoying and time consuming, on the other hand it is flexible as hell.

It still seems a bit overfit but I am just a bit tired of it.