Reference from Granblue Fantasy Character, Narmaya.
This Model is to achieve Narmaya Multi Version Costume that close to replicate her multiple attire in Granblue Fantasy with just 1 model by using specific trigger word, "ver1costume~ver8costume".
Summary about this model:
-Multi Version Costume are close represent of her Ingame variety attire
-Work on weight 1 to go
-Volatile, any of costume may mix up
-some checkpoint may end up less details
-ver1 and ver2 Costume highly volatile/easily mix due to similarities of both design.
===========================================================
How to use Lora:
LoRa:
-download the LoRa Model
-move this data into your Stable Diffusion -> Models -> Lora
LoCon & LoHa (LyCoris):
-Make sure your SD already contains LyCoris Extension https://github.com/KohakuBlueleaf/a1111-sd-webui-lycoris
-download the LoCon or LoHa models
-Move the data into your Stable Diffusion -> Models -> Lycoris
===========================================================
Trigger Words Simplify:
Default: ``1girl, narmaya \(granblue fantasy\), draph``
Optional: ``ver1costume~ver8costume`` and ``narunsfw``
Each of VerCostume Trigger will require ":1.2" regardless due to amount of different costume version and database been applied and current costume is easily volatile or mixing.
To prompt 1 of 8 VerCostume, below is an example:
Yukata Version:
``Yukata, (ver6costume:1.2), 1girl, narmaya \(granblue fantasy\), draph, <lora:NarmayaLoRa:1>``
Xmas Version:
``Xmas, (ver7costume:1.2), 1girl, narmaya \(granblue fantasy\), draph, <lora:NarmayaLoRa:1>``
Yukata Version:
``Butterfly, (ver5costume:1.2), 1girl, narmaya \(granblue fantasy\), draph, <lora:NarmayaLoRa:1>``
Maid Version:
``Maid, (ver8costume:1.2), 1girl, narmaya \(granblue fantasy\), draph, <lora:NarmayaLoRa:1>``
Non-VerCostume Trigger Words:
Other Costume that may have nothing to do with her ingame variation are also possible by only just use ``1girl, narmaya \(granblue fantasy\), draph, <lora:NarmayaLoRa:1>``. But the costume may reflect the same design among vercostume.
During self testing, it is found that certain VerCostume is unstable or over mixing which will be ver1costume and ver2costume since both sharing close to same attire but different layer clothes. This issues may resolve by next version.
===========================================================
This is my first time trying to get multiple unique attire of 1 character in 1 Model instead of describe every of their attire or outright does not replicate the same design. I be very apricate to get Feedback since I'm still continue self learning Lora Training.
If anyone wish to ask anything personally, directly contact or find me via CivitAI Discord by same user name, deatharms2010.
Cheers~
https://tensor.art/u/611028981541560770
Description
This is a LoHA model type of Narmaya Multi Version Costume.
FAQ
Comments (6)
Hi, really appreciate you trying to fit all her costumes into one. Curious question though, what would you say is the difference between each of the different types(LoRA, LoHA, and LoCon)?
Good question.
For your understanding, 3 of Lora been train with same parameters. The deference I found out from 3 of this is Quality / Source Learning / Files Size.
Quality
Both Lora and LoCon appears to have close to same quality, the difference is the LoCon is way flatten then LoRa. Both of Epoch are the same numbers, 20 Epoch.
LoHa on other hand have same flatten color pattern as LoCon probably way flatter but the huge different is the number of Epoch. As soon it pass over 10 Epoch, the sign of overcook it became to forming while other 2 LoRa able to fully complete the process with less possible overcooked.
**Source Learning**
Both LoRa and LoCon seems to have same level of abilities to interpret the image source while LoCon need way more or possibilities 2 times then other 2 Lora. You can see the example of image in About this Models, the LoCon Narmaya seem way lacking on interpret her left eye been cover by her hair.
Files Size
LoRa size over 30MB
LoCon size over 70MB
LoHa size over 100MB
Simplify:
If you want easy to go with any kind of Model training, it easy to say that LoRa is enough to cover all kind of form for your training.
Locon seems to be better at style, while loha is good at decontaminating styles during training
@mdebu5le yes it could be, every type of Dreambooth may have different purpose, that same goes for DyLora and iA3
@deatharms2010 Thanks for the really complete explanation!
Song/Tweyen bikini ver please!
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.