how to train embeddings just on earrings and clothes

[runner22k]

Can anyone help me with creating embeddings. I want to train 3 separate embedding on 3 subjects. (or is it possible to train on all 3, that would be awesome) I don't know how to tell Automatic1111 to choose only the dress from the photo and not jewelry, choose just earrings and nothing else.

1. Just earrings
2. Just the dress
3. just jewelleries

when we call dress, it should apply only dress type, and similarly with jewelries and earrings.

[Ackerlight]

An embedding is basically an embedded txt2img macro. It works completely on the txt2img layer and the description should be the opposite of how you normally create a description for a model (for example).

From personal experience, you may need to break this up into multiple embeddings. If you are all right with that, then describe every detail in the image except what you want. The embedding needs to infer (learn from what is not being described) what is missing from the images and it will recreate what is missing using txt2img.

Examples

1. A close up photo of a woman's ear. (AI will focus on the ears and see you did not describe the shiny things hanging off them. So will learn to recreate earrings from what you did not describe).
2. A woman with long blond hair standing in a room with hardwood floor, white walls and white curtains and wearing high heels. (AI will see the woman and notice you did not describe what clothing she is wearing. So may begin to look at what is similar between photos and notice her dress).
3. A close up photo of a woman's neck. (AI may notice here necklace).
   A close up photo of a woman's wrist. (AI will look at her wrist and see you did not describe the watch or bracelet. So will begin to look for similar bracelet and recreate it).

Now if you instead say "A woman wearing a blue sapphire princess cut necklace on her neck" the txt2img macro will ignore the necklace and never search or learn about it. Because you already included it in the txt2img prompt. Anything you include in the prompt will not be included in the embedding and the embedding will assume the user will need to type that every time as part of the text. It will be completely ignored for training.

[minienglish1]

That's great to know. I've read several papers and blogs about embeddings multiple times and I still didn't realize it worked like that.

Where did you learn that? Do you know where I can find an information source that covers other related concepts in such detail?

Thanks.

[Ackerlight]

Mostly from the SD Discord and other sources on the basics. The rest is from personal experience having created about 80 or so embeddings and going through trial and error.

[minienglish1]

Thanks.

[andupotorac]

Have you created any embeddings of clothes you're willing to share so we can test how they work? Will they work on any character - be it a person, or an animal in human form for example?

[andupotorac]

Wanted to follow-up on this. Yes, they work on any character, and style it seems. I just trained a few clothes.

A cap

[runner22k]

:) Thank you very much for a detailed description @Ackerlight . Sorry for delayed reply. I didn't had access to system from few days.