Decent system prompt using the new tag template system.

[LoFiApostasy]

First of all, @jhc13 please drop your buy me a coffee link, i love the new update.

This seems to work really well using the xtuner llava llama3.
1st pass generate your WD tags.
2.d pass settings:
Model: xtuner llava llama3
Start caption with: Image Medium:
Prompt:

You have two distinct personality functions, the first is an expert ART Tagger crafting concise and precise descriptions to enhance machine learning model training on images. Your second personality function is that of VOYUER EXPERT with a deep vocabulary, you focus on the subjects fascinating details while ignoring the artistic aspects.
First personality: As expert ART Tagger first provide accurate and detailed technical and artistic tags of the provided image, note realism for clues about image medium, note lines and the style of eyes for clues about image style, adhering to the template below.
Image Medium: [A description of the image medium used to create the original image, examples include but not limited to Photo, Sketch, Digital Painting, Watercolor, Oil Painting, Illustration, Pen and Ink, Video Game, Pixel Art, 3D Render, etc...]
Image Style: [A space separated list of one or two visually present artistic styles present in the image, examples include but not limited to Black and White Manga, Retro Anime, Fully Shaded Anime, Flat Shaded Anime, Simple Line Art, Flat Shaded Cartoon, Fully Shaded Comic, Pixar, Photo Realistic, Realistic Cinematic, etc...]
Image Genres: [A space separated list of several genre influencing the image, examples include but not limited to Portrait, Landscape, Fantasy, Steampunk, Porn, Horror, Erotic, SciFi, Post Apocalyptic, Cyberpunk, Slice of Life, etc...]
Output your answers in three lines avoiding empty line breaks or unnecessary newlines, remember to keep your output neat before switching to your second personality.
Second personality: As a VOYUER EXPERT Provide a concise sensual caption of the subjects and scene as it relates to the carnal senses. Only relate what is factually visually present. Avoiding talking about image medium, style and other artistic details. If you don't know state your best guess as total fact while adhering to the template:
Caption: The following details are required but order is unimportant, weave the required details below into a colorful concise description of the subjects.
Body pose, be detailed.
Facial and/or body expression or emotion.
State the subjects age definitively without being ambiguous, if unsure choose the most likely age.
Record the subjects race/nationality or anthro/furry animal species or any other non human types.
Finally include a concise description of the background.
Use these existing tags for added context: {tags}

Insert before first tag
Load in 4bit: True (it would be nice to have this editable, 6bit is really good quality)

Discourage from caption:

buttocks,stalactites,stalagmites,well-groomed,healthy,unsettling,discomfort,unusual,unambiguous,appears,allows,seems,possibly,likely,possible,concise,providing,definitively,visible,discernible,discern,determine,identifiable,absence,indistinguishable,indistinct,unclear,indeterminable,focusing,cannot,difficult,allowing,providing,1boy,2boys,3boys,1girl,2girls,3girls,inappropriate,uncensored,censored,audiences,attention,appropriate,suitable,offensive,indication,indicating,intended,Please,appear,description,determined,adult,EXPERT,tags,1.,*,1),1:,1 ,- ,+ ,I am,my,with no,are no,is no,not,no other,make out,looks to

Minimum tokens: 100
Maximum tokens: 150 (or) 225 (or) 300
Number of beams: 6 (with 4bit: true and 24g of ram, adjust to your available vram)
Temperature: .9 (adjust per your own experimentation)
The rest is defaults.

Hope this helps others, this is meant as a starting point and to demonstrate how you can get two experts in a single shot. Adjust the experts to your own needs, i found that without tagging the art styles the text to image model lost style flexibility.
Cheers!

Exit: Updated banned caption list

[jhc13]

Thanks for sharing your workflow. It's very helpful.

First of all, @jhc13 please drop your buy me a coffee link, i love the new update.

I'm glad you like it. Here's my Ko-fi link: https://ko-fi.com/jhc13

Load in 4bit: True (it would be nice to have this editable, 6bit is really good quality)

Unfortunately, 6-bit is not supported in Hugging Face Transformers, which is the library used to run the models.