phantom-diffusion-s3-the-last-8

The final 8 models that are trained over hakurei's Waifu Diffusion.

Each model was trained on a notable Japanese (and Taiwanese, I think) AI artist's works using dreambooth, with 30 of their works gained mainly from twitter (except for sabakichi, which I collected the dataset from pixiv).

It tooks 1000 steps to train each model, at a learning rate of 5-e6. I have shared the notebook here in case you want to try out additional training.

Some are good at backgrounds, while others can generate a semi-realistic style as seen in many SD 1.5 anime/waifu models (and some might be a bit overfitted).

The dataset is published here, and the full members are listed here (special thanks to Yunyalula for creating it!).

For those who are against generative AIs

You can see that now they are the target. Take our tool. Now is the time for pay back. Generate the images in their styles, and bring back the power you had to yourself.

For those who support the development of generative AIs

Some of the AI artists, even though they take advantage of the open strategy of Stable Diffusion, now tend to hide their prompts, trying to monopolize their style (I'm not saying the AI artists I trained are as such, to be sure).

To continue protecting our values and beliefs on the open community and fight against them trying to create another pre-modern style guilds, I will show you a new way.

You no longer need their prompts; just train their images by yourself to protect the open community. It's not only legal but also ethical, as they have been taking advantages of others' trained dataset.

For those who call themselves "phantom 40"

I saw some caliming there should be 48, and here you go. Phantom 48, or do you like to call yourselves PTM 48 instead? It's up to you.

Why will they be the last?

My initial intention on this series was a social experiment to see what will happen if the AI artists are targeted for personalized training.

As it became more popular than expected and the artists started calling themselves "phantom 20," I came up with the second intention to see how they will react after I add 20 more in one day, to see if they can adapt to the sudden change. They acted greatly, and I think that's why they could become notable.

All the reactions and the interpretations on my action were impressive, but since I have accomplished my goal, and since the main stream model will probably be SD 2.1 768 (not SD 2.1 512), I will no longer add new models.

I know I couldn't add some of the artists, but no. I will not do it under the name of phantom.

It takes me like 8 hours to train, test, and upload 20 models, and it's just unsustainable to continue doing it everyday.

From now on, anyone who wish to add more is the next phantom. Train anyone you wish to by yourself.

trained artist list

• atsuwo_AI
  • recommended pos: multicolored hair, cg
• fladdict
  • recommended pos: oil painting/ancient relief/impressionist impasto oil painting (maybe more)
  • possible neg: monkey
• Hifumi_AID
  • recommended pos: dark purple hair, emerald eyes
• mayonaka_rr
  • recommended pos: cg
  • possible pos: dynamic posing, bikini, ponytail
• o81morimori
  • possible pos: cg, in a messy apartment room with objects on the floor and the bed
• sabakichi
  • possible pos 1: merging underwater, limited pallete, melting underwater, unstable outlines
  • possible pos 2: rough sketch, limited pallete, ((unstable outlines)), monotone gradation, dynamic posing
• teftef
  • possible pos: light skyblue hair, bun, retropunk gears of a factory
• violet_fizz
  • recommended pos: beautiful face, grown up face, long eyes, expressionless
  • possible pos: expressionless

samples

The basic prompt is as follows.

However, to present you the potential of these models as much as possible, many of them have additional postive tags (such as "in the style of") to get the result below (yes, use aitop (ARTIST)_style to gain the finetuned result).

Many works better with the additional prompt beautiful face. Generally speaking, prompting words close to the trained dataset will give you a better result.

POS: masterpiece, best quality, 1girl, aitop (ARTIST)_style
NEG: nsfw, worst quality, low quality, medium quality, deleted, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digits, fewer digits, cropped, jpeg artifacts, signature, watermark, username, blurry, simple background

atsuwo_AI

fladdict

Hifumi_AID

mayonaka_rr

o81morimori

sabakichi

teftef

violet_fizz