Type | |
Stats | 3,880 635,491 |
Reviews | (230) |
Published | Oct 30, 2023 |
Base Model | |
Training | Steps: 99 Epochs: 3 |
Trigger Words | ca45mv7-100 |
Hash | AutoV2 3C1092CB7F |
Disclaimer: Any resemblance with any person living or dead is pure coincidence.
Notice for Civitai generation service users: When using Civitai generation service with embeddings, you still have to manually copy and paste trigger word in your prompt (or in negative for negative embeddings). Otherwise your generated images won't be affected by embedding model. On how to best position character embedding within the prompt, please refer to Usage tips i have provided in description of the model down below. Also, model trained for SD 1.5 and not compatible with SDXL checkpoints.
I have converted .pt into .safetensors for your convinience. Usage is same as with .pt. Both file types available for download.
UPD:
added Variant 3 (Jaka).
added Variant 2 (Sofi).
Variant 1 (Maja):
Face. High consistency, will vary according to SD model associations, you can freely change hair color, hairstyle, and so on; some features may require additional (emphasis) to be changed.
Body shape and height. High consistency.
Breast size and shape. High consistency.
Spirit of sexiness. This TI has been trained on nudity. If you want to produce SFW images use clothes in your prompt or "nude" tag in negative. Unlike my first model (vol.1) this time it's not overtrained, usually it's not necessary to apply high weights for clothes.
Vectors per token: 7. This means that embedding will keep most of the character appearance approximately within first 50 tokens of your prompt. This also means that embedding have moderate token count and may interfere with your prompt, if it have complex scene that not registered within embedding (to deal with it - see usage tips).
Variant 2 (Sofi):
Face. High consistency, will vary according to SD model associations, you can freely change hair color, hairstyle, and so on; some features may require additional (emphasis) to be changed.
Body shape and height. High consistency.
Breast size and shape. Moderate consistency.
Spirit of sexiness. This TI has been trained on nudity. If you want to produce SFW images use clothes in your prompt or "nude" tag in negative. Usually it's not necessary to apply high weights for clothes.
Vectors per token: 6. This means that embedding will keep most of the character appearance approximately within first 30 tokens of your prompt. This also means that embedding have moderate token count and may interfere with your prompt, if it have complex scene that not registered within embedding (to deal with it - see usage tips).
Variant 3 (Jaka):
Face. High consistency, will vary according to SD model associations, you can freely change hair color, hairstyle, and so on; some features may require additional (emphasis) to be changed.
Body shape and height. High consistency.
Breast size and shape. High consistency.
Spirit of sexiness. This TI has been trained on nudity. If you want to produce SFW images use clothes in your prompt or "nude" tag in negative. Usually it's not necessary to apply high weights for clothes.
Vectors per token: 6. This means that embedding will keep most of the character appearance approximately within first 30 tokens of your prompt. This also means that embedding have moderate token count and may interfere with your prompt, if it have complex scene that not registered within embedding (to deal with it - see usage tips).
All features of all versions will vary depending on the SD model you use.
Usage tips:
Basically vectors per token parameter of character embeddings determine how much info can be learned about subject from training dataset you provide (higher = more info; it is mostly depth maps and some characters cannot be learned without sufficient vector count). Downside of high values for vectors per token is loose of flexibility within prompt (embedding will resist injection of anything that wasn't learned during training). Fortunately there is 2 methods to gain flexibility, and if you feel that embedding not allow you to get results you wanted use one of those:
Method 1. Positioning or weight adjastment (works with Civitai generation service).
You can either move embedding farther from beginning of the prompt (e.g.: RAW photo of woman, your lengthy prompt, embedding_name, your prompt again), or lower weight of embedding (e.g.: (embedding_name:0.8)). Both variants works identicaly - flexibility raising while likeness decreases. Raising weight of character embeddings will not improve likeness, but may add some details or introduce you some artifacts.
For Civitai generation service users:
Service currently is very raw. You have no choice but to position embeddings somewhere close to the beginning of the prompt (e.g.: RAW photo of woman, embedding_name). Note that moving embedding too far away from the beginning will completely "turn it off". How far depends solely on Vectors per token parameter with which it was trained (see above). Since service don't have even simple token counter you have to experimentally look for a good spot.
Method 2. BREAK word (doesn't work with Civitai generation service).
You can type all you want in your prompt, and then at the end add word BREAK (must be uppercase) after which add character embedding (e.g.: RAW photo of woman, your lengthy prompt BREAK embedding_name). This way you will get flexibility according to actual position of embedding in the prompt (farther from the beginning = better) while all likeness will remain intact. Most example images has been made using this method.
If you like my work, please consider to throw some stars at me. Also i would appreciate any feedback on the model and posts of your work with my embedding.