Skip to main content

Character Consistency

Feature Requests

Why is there no way to create character consistency? Especially or at least facially. Or am I not seeing where to do that? Also no photo upload for face swap or to create character prototype with photo and place them in different settings and clothing.

Status: Rejected3 comments

Log in to comment and vote

Comments3

  • Blue Storm

    •

    Mar 6, 2025

    Some other suggestion (some of which follow the blog suggestions). It’s never perfect, but it help.

    • Use danbooru tags, which lower the impact of fluent language conversion to instructions. Fix the seed in the image generation parameter.

    • Split clearly the character definition and the pose / context. This would help to get similar results.

    • Fix the image seed in image context


    But yes a mecanism allowing to pass a reference image would be awesome.

  • Teal Lightbulb

    •

    Feb 9, 2025

    For me, The way I set the prompt up helps me generate a consistent character style and look. Here is an example of a prompt template that I use when creating character. The word weighting you see below is fairly straight forward. The main things to keep in mind are:

    0.0 thru 2.0 are your normal strengths. 0.0 will be ignored, meaning a word that is formatted like (red:0.0) will have no affect but (red:0.1) will have a slight affect. You should avoid forcing anything over 2.0 as strengths higher than that will typically break the image but not always.

    You can also create a prompt like my example below that has several options listed that you can turn on/off and.or mix to get different results with the same character but sometimes you will need to regenerate the image a couple of times to get it back on track but not always. an example of this would be:

    He is wearing (red:0.0) (blue:0.0) (Green:0.0) (black:0.0) gym shorts and a (red:0.0) (blue:0.0) (Green:0.0) (black:0.0) pull-over hoodie. which ever color you “turned on” by changing the weighted number (0.0 to 0.1 thru 2.0) will have more weight with the generator and usually be the color that is rendered. can also adjust the strengths of primary colors to achieve intermediate colors.

    This is something you can do for almost every adjective and verb in your prompt. As long as there is a logical choice you can triggering/blend it with weighted words/phrases.

    Create a highly detailed full body image of a (age, gender). (He/She) has (skin color) skin and a (Facial characteristics:1.0). His/Her eyes are (shape) with (Color) eyes. His/Her mouth is (open/closed/smirk etc.).
     
    (He/She has perfect anatomy with no distorted or facial deformities, (lip shape):1.0)
     
    (length and color) hair (what is the hair doing?) over his/her (face, back, shoulders). He/She has (bust/chest size:.1.0) with (body type (i.e. athletic, slender, slim, obese etc.:1.0).
     
    wearing a (style of clothing, including fabric and ;
    Soft gradient background, evenly lit. (rim lighting, studio lighting, dslr, ultra quality, sharp focus, tack sharp, dof, film grain, Fujifilm XT3, crystal clear, 8K UHD) ß use this section for the camera and to set up the shot.

  • Silver River

    •

    Feb 8, 2025

    The Blog has some articles on how to do that by using consistent seeds and prompts and that could help you. More advanced ways of doing it aren’t implemented yet, though I think they should allow you to upload images to train a LyCORIS or something for a character. Then you could upload the LyCORIS to memory when you want to generate something with that input art style, character etc as they’re usually pretty small in size.