prompt help - questions, ideas, experiments...

linvanchenelinvanchene Posts: 1,382
edited May 9 in Daz AI Studio

Learning how to put together prompts in Daz AI Studio is like learning a new language.

The idea of this thread is to help each other out when we are stuck with our prompts.

 

You have a great idea for an image in your mind?

You need help to phrase it in a way the AI Image Generator understands?

Ask away. Someone on this forum may have an idea.

Anyone can ask and help out...

 

General tips:

For starters check this blog entry:

Daz AI Studio: Prompt Engineering Guide - Daz 3D Blog

Are you looking for existing prompt examples? 

Curated lists by Bunyip02:

https://www.daz3d.com/forums/discussion/683016/are-you-looking-for-prompts

 

- Describe your basic image idea with as few words as possible

- add one more word and observe how the image changes

- combine a subject with two or three additional objects, a background, a lighting style

- the more words you use to describe your image the less effect each word will have

- words at the beginning of the prompt are usually of higher priority

 

 

Post edited by linvanchene on

Comments

  • linvanchenelinvanchene Posts: 1,382
    edited May 8

    How do you make animals look like toons?

     

    close-up view of a  squirrel, oak leaves, forest, 

     

    Add the one word that changes the style

    + toon

    close-up view of a toon squirrel, oak leaves, forest,

     

    Add additional words that reenforce the style

    + fluffy 

    close-up view of a fluffy toon squirrel, oak leaves, forest,

     

     + cute

    close-up view of a cute fluffy toon squirrel, oak leaves, forest,

     

    Using words that are synonyms can strengthen the theme.

    At some point adding more words that affect the same area will not anymore make a huge difference and may have unwanted side effects.

    One could argue that the squirrel may look more cute. But it also looks less fluffy now...

    I ususally don't use more than 3 attributes for one object.

     

    Mix similar animal shapes together

    You can change different parts of the body by morphing them into a feature from another animal.

    + bunny ears

    close-up view of a toon squirrel with bunny ears, oak leaves, forest,

    This works best if the base shape of the animals is similar...

    It is a matter of trial and error to stumble upon combinations that work...

    dazai_51f7be8c-0d8b-11ef-ba0f-e9c8be489245.png
    896 x 1152 - 1M
    dazai_845942ba-0d8b-11ef-ba0f-e9c8be489245.png
    896 x 1152 - 1M
    dazai_c08fca06-0d8b-11ef-ba0f-e9c8be489245.png
    896 x 1152 - 1M
    dazai_04c8eb58-0d8c-11ef-ba0f-e9c8be489245.png
    896 x 1152 - 1M
    dazai_5441647a-0d8d-11ef-a784-a702dcdfccf3.png
    896 x 1152 - 1M
    Post edited by linvanchene on
  • Nyghtfall3DNyghtfall3D Posts: 776

    linvanchene said:

    - the more words you use to describe your image the less effect each word will have

    This part is constantly throwing me, whether Daz or another app.  I keep wanting to know if there's a cutoff, and if so, where it is.

  • linvanchenelinvanchene Posts: 1,382
    edited May 9

    Rephrased several times. Hope this version makes sense. blush 

    Nyghtfall3D said:

    linvanchene said:

    - the more words you use to describe your image the less effect each word will have

    This part is constantly throwing me, whether Daz or another app.  I keep wanting to know if there's a cutoff, and if so, where it is.

     

    Cut-off  point 

    Some guides talk about a limit of  tokens after which inputs are not anymore evaluated.

    tokens: words, punctuation marks, symbols, brackets

    Anything written in the prompt that needs to be evaluated...

    That token limit for SDXL is indicated by others at ~ 75 tokens.

    (Never counted it myself. )

     

    Words compete for dominance in the same area

     

    Some words may affect the whole image. Other words may only  affect a small area.

    Words that affect the same area "compete" with each other.

    Place the words that should have priority first.

     

    The longer the prompt gets the higher is the chance that a word at the beginning of the prompt already "claimed" an area of the image.

    The noise in that area is not anymore reshaped and the word has no effect even though the cut-off point was not yet reached.

    In the current version of Daz AI Studio this noise reshaping is not shown in the user interface.

    All users can observe is a spinning circle.

    The process looks similar to an Iray render starting from pixelated noise but during the generation process some areas are reshaped in random ways.

    Examples:

    jacket,  shirt

    Both the jacket and the shirt occupy the same area.

    Sometimes a closed jacket is generated and the shirt is not visible.

    If you place the shirt first sometimes the jacket may not be generated.

     

    jacket, pants

    Word order does not matter because object areas do not overlap.

    Post edited by linvanchene on
  • FirstBastionFirstBastion Posts: 7,760

    Absolutely,  the words before the first comma are generally the most influencial.

  • linvanchenelinvanchene Posts: 1,382
    edited May 9

    Bunyip02 said:

    https://www.daz3d.com/forums/discussion/683016/are-you-looking-for-prompts

    Thanks Bunyip02. I added your link to the first post.

    Curated lists of prompts are a great starting point to experiment.

    Add your own words, observe what changes.

    Remove some words to check if they AI Generator recognizes them and alters the image in your imagined direction.

    Post edited by linvanchene on
  • FirstBastionFirstBastion Posts: 7,760

    There does seem to be a word limit that is allowed into the prompt field here at Dazai. 

  • linvanchenelinvanchene Posts: 1,382
    edited May 23

    Can you write prompts in Daz AI Studio in other languages?

    I just entered some words in german and was surprised to see matching images generated.

     

    A quick google search showed that during the SDXL base model training some images may also have been captioned in german, french, spanish and other frequently used languages.

    Based on that prompting in different languages may work for common concepts but not yield the expected results with more advanced descriptions.

    daz-ai_user-interface_overview_german.jpg
    1143 x 769 - 195K
    Post edited by linvanchene on
  • FirstBastionFirstBastion Posts: 7,760

    Given that some of the first functions for AI was spell check and translations,  it makes sense it understands various languages  and even mispelt words.

  • linvanchenelinvanchene Posts: 1,382
    edited May 27

    Why does the Angela 9 LoRA sometimes generate white women?

     

    LoRA usually are not just an on - off switch.

    To activate them you also need to add the triggers to the prompt - tokens, words that were used in training.

    The base triggers for character LoRA are

    • man
    • woman

     

    In case of Angela 9  an additional trigger can be "dark skin".

     

    Add tokens like "dark skin", "tanned", "tanlines", "light skin", "pale skin",  to get different skin tones.

     

    A different character in the same LoRA with another trigger?

    In Daz 3D Studio it is possible to apply a different texture to the Angela 9 shape.

    I am not sure if what we see here is a version of the Angela 9 shape with a white skin texture.

    In that case the Angela LoRA could have been trained with a second trigger like "pale skin"

     

    a default checkpoint face

    Or - this could just be one of the "default checkpoint faces" you get when you enter certain token combinations in the prompt.

    In any case, that face triggers consistently when you combine Angela 9 with "pale skin".

     

    What happens if we do not use trigger words in the prompt?

    • Sometimes you do not need to add trigger words. The intended effect will just trigger by activating the LoRA at a specific strength.
    • Sometimes the LoRA will just randomly select combinations of the trained triggers if they are only partially triggered.
    • In other cases without the proper trigger words nothing happens.

     

    Usually you can figure out  the trigger words by looking at the LoRA training data.

    It could be that some of the character LoRA have some alternate trained shapes that you could trigger with the proper combinations of words.

    adult, younger,  curvy, model, muscular...

     

    ###

    The developers who create Daz AI Studio may be able to provide some background information on such questions - to prevent speculation.

    daz-ai_angela-9_skintone.jpg
    1038 x 750 - 149K
    Post edited by linvanchene on
  • linvanchenelinvanchene Posts: 1,382
    edited May 26

    How can I make characters look older?

    You can make characters look older by adding "old woman" to the prompt.

     

    You can also get more specific by adding an age - "30 year old woman", "70 year old woman".

     

    portrait of a 70 year old woman, lob hair, dark skin, white blouse, glasses, reading a technical manual, computer, office, bokeh,

     

    Edited: I split this up into two posts - may be easier to link when the same question shows up again later. 

    daz-ai_angela-9_old-woman_glasses.jpg
    1048 x 357 - 77K
    daz-ai_angela-9_xy-year-old-woman.jpg
    1038 x 799 - 160K
    Post edited by linvanchene on
  • FirstBastionFirstBastion Posts: 7,760

    You have made an interesting discovery linvanchene. Why there is variation in the influence is worth further investigation.

    I used your prompt with the "pale skin" keeping all other things equal,   and each character's LoRA  influenced in various ways the output, except for  Angela, and perhaps to a lesser extent Wang. 

     

  • linvanchenelinvanchene Posts: 1,382
    edited May 27

    That is an amazing overview, First Bastion - it illustrates that you can get feminine versions of male LoRA by using the token "woman".

    This made me try this in reverse with "man".

    kept the prompt the same:

    portrait of a man, bob hair, masculine, close-up of face, white blouse, park, pond, noon, bokeh

     

     

    masculine - feminine ?

    Edited after trying it again.

    I was not able to replicate consistent results when trying to "gender-swap" the LoRA.

    When using the token "man" with Kala 9 sometimes still a feminine version would be generated. etc.

     

    • man
    • masculine
    • man, masculine
    • woman
    • feminine
    • woman, feminine

    The tokens man, woman worked best when trying to trigger the LoRA.

    I thought using "man, masucline" and "woman, feminine" would reenforce the theme. But that did not work consistently.

    Using just the tokens "masculine", "feminine" did not seem to have a consistent effect on its own.

    We probably would need some feedback from the developers if masculine and feminine versions for each LoRA are

    • an intended feature - trained with images, tags and triggers
    • just a side effect of how generative AI works...?
    daz-ai_man_overview_2048x5176_mq.jpg
    2048 x 5176 - 776K
    Post edited by linvanchene on
  • linvanchenelinvanchene Posts: 1,382
    edited May 27

    What are styles?

    A style is a token that influences the image generation in a massive way.

    It can have effects on the color scheme or on the shape of objects.

     

    Genre

    You can find some examples in this blog entry:

    Daz AI Studio: Prompt Engineering Guide - Daz 3D Blog

    • fantasy
    • gothic
    • science fiction
    • post-apocalyptic
    • steampunk

    Just adding that one word can influence the clothing, shapes, buildings in the background.

    Other examples:

    Synthwave

    synthwave, photo of an iridescent sports car on a highway, waves, palms, desert, sun, moon, motion blur,

    This prompt includes other tokens that influence the image

    • iridescent - color changes depending on surface angle
    • motion blur - does not seem to have a consistent effect, seems to happen randomly with moving objects

     

    Isometric

    Can be a very  specific style on its own. In Daz AI Studio it mostly shifts the perspective similar to "view from above".

    isometric, city street,

     

    If you search the internet for "styles for Stable Diffusion XL" you may find plenty of inspiration.

    Which are your favorite styles for Daz AI Studio?

     

    daz-ai_styles_synthwave.jpg
    1300 x 439 - 117K
    daz-ai_styles_isometric.jpg
    1299 x 320 - 149K
    Post edited by linvanchene on
  • linvanchenelinvanchene Posts: 1,382
    edited June 8

    xy driving a [enter vehicle name here]

     

    I looked up the prompt I used on those photos with a driver in a car - you may need a lot of rerolls to get a useful one.

    Depth ControlNet would be extremely helpful at this point to block in the base pose.

     

    for a close-up:

    photo of a woman driving a sports car, close-up of face, highway,

     

    for long shot:

    photo of a woman  driving a sports car, isometric, highway,

    photo of a woman  driving a sports car, view from above, highway, 

    Post edited by linvanchene on
  • linvanchene said:

    Mix similar animal shapes together

    You can change different parts of the body by morphing them into a feature from another animal.

    + bunny ears

    close-up view of a toon squirrel with bunny ears, oak leaves, forest,

    This works best if the base shape of the animals is similar...

    It is a matter of trial and error to stumble upon combinations that work...

     Hey! A springhare! 

Sign In or Register to comment.