Announcement - July Changelog (Patreon)

Hey Jibaku here º^º/

General Updates

• The poll has finish and we have a winner !!!

• Introducing new tags with for July Monthly Style (MS0724)

• Introduction of Fallbacks Styles

• Make difference between SD style and SaA Style (Same as Anime)

• Increase MAX number of images in one set from 40 to 60. More information in the Fallback graph

Updates about Twitter

I think that i am popular with Pixiv but i don't really think that i am popular on Twitter.

So i am looking for ideas to attract more people and grow my Twitter a little bit.

If you have any idea, you can helping me adding it as part of the comments in this posts.

A few ideas that i have to grown Twitter are:

• Add one SFW picture of the current set that will not be feature in Pixiv

• Add one NSFW uncensored picture of the current set that will no the the same as Pixiv

• Use more Hashtags in posts

• Make half-month poll and create a free set with Monthly Styles

• Open a new account since i already follow a lot of people and i suspect that Twitter Algorithm don't like that i have more following people rather than followers

Updates about sets and Fallback Styles

Problematic

As you may notices, the past month there was a set with Firefly from Honkai Impact.

( https://www.patreon.com/posts/firefly-honkai-106526888 )

This set is different from the other sets because the way of generation differ from the regular sets.

Originally the software used to generate the images only contains characters featured from 2023 or less. This means that new characters from 2024 like Frieren, Ubel, Firefly or Kikoru (Kaiju 8) are not part on the original training set and the tag of the character to generate the image is missing. Also the Software is privative and does not allow create new tags or feed the network using new images.

Returning with Firefly, Since i wanted to post about Firefly. I had two options:

1.- Generate images using only the tags to simulate the characters but it will lack of details

2.- Generate the images using SD+Lora and then apply a filter to simulate the target style

The issue with No. 1 is that a lot of details are lost and something didn't really looks like the character.
The issue with No. 2 is that sometimes the filter style and the original content of SD fighted causing that the image chroma gets worst (CFG Burn). Also the time of generation was very huge, around 5 minutes per image.

Solution

In order to complete all the requested character, i introduced a "fallback" style.

What does this means? That you can request a character and in case that could not be generated using the Software, i can apply the same techniques used in the Firefly Set Option 2.

But the final decision will be you, if you decide to use the Fallback Style or request another character with the native style that you wanted.

You can check the chart in the Attachments section for more details about the Fallbacks Styles.

The graph contains the name of the Style, [The number of pictures that i can generate using that style] and (more explanation about the style).

Sorry if this gets more complicated that expected ....

In case of any question, you know that you can reach me by DM. \º^º/

 Thanks for read and continue the support until now