Transitioning to V5 from earlier versions

Transitioning to V5 might require some tweaking of your setup if you use custom avatars. V5 now relies more on avatar descriptions than avatar image. This applies most to photoreal, but similar concepts apply to anime as well. First, try to get close using avatar boosts in custom avatar. This is the best way to achieve consistent results if you're coming from previous versions - use at least 4 images from previous versions you like, and apply it at moderate weights as a starting point. Avatar description should include things like skin tone, hair style and color, face shape, body features. Write them in natural language phrases instead of comma separated keywords for best results. Use the wand tool in custom avatar to get a start and refine its description. Writing the avatar description in comma separated list form is the biggest culprit to avatars not looking exactly right - ask the community for help here if you need to!

The default face detail in V5 is now 0%, down from 50% in V4. This means you can start from 0 to have maximum fidelity to the identity in the avatar photo - then increase if you want more detail like brightly colored eyes or freckles. Avatars in V5 will look similar for the most part but distinctly different than what you may used to due to rendering style - you don't have to switch to V5 if you don't want to, and you can take your time on V4 or V3 as you explore switching. All V4 models are still be available in model select.

V5 Prompting

The V5 image generation system uses natural language prompting instead of the keyword-based approach of previous versions. While earlier versions used syntax like (concept:1.2), V5 prefers narrative descriptions that paint a scene. In addition, avatar description is incredibly important in V5 so see the appropriate section for tips on getting it right.

Prompting Enhancements

Note: Weighing, pose reference, and style reference sections are relevant only to V5 and earlier models; the latest engine no longer uses these systems.

Weighing

You can further control a prompt by using weights, which modify the strength of certain concepts. To weigh a subject higher, enclose it in parentheses and specify a weight, like (blue eyes: 0.5). This would weigh "blue eyes" at half the impact it would normally have. To emphasize something more, use (broad shoulders: 1.2) to emphasize it more.

Pose reference

You can upload (or paste in) an image to be the pose reference for a given prompt. Pose reference tries to match the output image as close to the given pose as much as possible. With pose reference, even the body shape of the main subject(s) will be taken into consideration in the output to match close to the reference image. Pose references are the most powerful way to control the composition of the output.

Style reference

You can upload (or paste in) an image to be the style reference for a given prompt. Style reference is a visual way to influence the final image, or to have the final image incorporate elements from the style reference.

Using the Prompt Expansion Wand

• Click the wand icon to automatically expand simple prompts into more elaborate, descriptive ones that work better with the image generator
• This helps create natural language descriptions while maintaining all necessary details

Effective Prompt Elements

A good prompt doesn't need to be long but should paint a clear picture. Key elements to include:

• What they're doing
• What they're wearing
• Where they are
• Environmental details

Even just a sentence or two covering these elements can create a full scene.

Elements to Avoid

• Specific poses (use pose reference instead)
• Poses like "lying down" are prone to limb mutations and deformities
• While you can describe visual styles, avoid using "in the style of <artist name>"
  • Instead, use descriptive words to convey the desired style
  • Style descriptions can be effective at high weights (3-6)
• Negative prompting:
  • V5 doesn't use /// syntax. Things after /// will be ignored
  • Use natural language like "avoid x" instead, which can work conditionally for certain concepts. Best to describe what should exist in the image than what should not

Weighing System in V5

• Syntax: (subject or phrase: weight), like (subject or phrase: 1.5)
• Weights can go higher than previous versions:
  • 4-6 can be ok for very strong emphasis
  • What used to be 1.2-1.3 for features can now be 1.5 or higher
  • High weights on avatar features for individual Kindroids can cause bleeding with others in group selfies - don't go overboard
  • Going over 7-8 is unlikely to help and may make the image rigid
• Best Practices:
  • Weight larger parts of the prompt rather than specific words
  • Entire phrases can be weighted: (Something elaborate and detailed: 1.8)
  • For anime Kindroids, use high weights to get strong emotions and facial expressions
  • In group selfies, high weights can bleed into other Kindroids - find the lowest necessary weight that gives consistent effects without bleeding

Avatar Configuration

Custom Avatar Description

• V5's version of avatar fidelity is avatar boost - find it in the bottom of custom avatars, and it will be very helpful as a starting point for tweaking further
• Avatar description matters more than ever in V5 and is very important for the AI to render your custom avatar correctly
• Keep descriptions narrative - expand on keywords. Length does not need to be long as long as it’s clear in a narrative form to a human (or AI)
• Use the wand or AI to describe the avatar image as a starting point
• Include:
  • Skin tone
  • Gender
  • Ethnicity
  • Hairstyle and color
  • Body shape
  • Special features (tattoos, etc.)
  • (Note - eye colors won’t be salient when put in avatar description. They belong better in face prompt - see below)
• Weight important features higher but use weights above 2 sparingly
• For anime style, including a style with weights 2-3 can help lock in a preferred visual style

Face Detail Slider & Prompt (Photoreal V5 Only)

• Default: 0 (highest likeliness with avatar image face)
• Higher values:
  • More detailed faces
  • May deviate from avatar image
  • Required for specific face features (freckles, bright eye colors)
  • Use with face prompt for best results
• Eye colors and other face features belong better in face detail prompt, coupled with a moderate to high face detail value. 
• Face prompt specifically targets facial features, whereas avatar description covers general appearance
• Face prompt guidelines:
  • Best when face detail setting ≥ 30%
  • Uses a keyword system (not narrative), and weights should NOT be over 1.3 to 1.4. Face detail prompts use a different prompting system than V5 in general. 
  • Example: (freckles: 1.2)
• Face prompt only works in photoreal V5, though face detail slider applies to legacy models

Common Issues and Solutions

Photoreal

• Females - Unwanted large breasts:
  • Describe clothes: (wears modest and conservative clothes: 2)
  • "She is a small stature (petite woman with a minimal and natural chest:1.3)"
  • "She is small-framed, delicately built and has subtle natural curves"
• Males - Reducing facial hair:
  • Use terms like: "beardless", "clean shave", "clean-shaven"
    • These words in avatar description will cause male photoreal avatars to have far fewer cases of strong facial hair through a patch that detects the above keywords
  • Use terms like “young” to create a young appearance. Note that “young” does NOT apply the facial hair patch, but it works effectively to reduce facial hair. You can then use “mature, 30 years old” in face detail to age the face up if necessary, as an alternate way to remove facial hair without the facial hair patch. 
  • Special keyword "zerobeard" in avatar description will harshly remove beards, but cause avatars to become noticeably younger
• Males - Reducing chest hair:
  • In avatar description: (hairless chest:2)
• Males - longer hair
  • Community contributed unofficial tests & guide: Guide
• Bobblehead/animated glitch fixes:
  • Terms such as: "porcelain", "japanese", "pixie", "doll", "petite", "bobbed", "anime", "pixar", "stylish" have been flagged as higher risk for turning faces into animated versions.
    • These terms, when detected, will result in the system re-emphasizing photoreal - this should not affect much but this is a system level prevention against bobbleheads. If you continue to run into the issue, lower the weights or remove words like these. 
  • Other terms may cause photoreal to turn stylized and cause face distortions too. Fantasy elements and others should use your own experimentation and adjustment on weights. 
• Long neck issues:
  • Be cautious with thin descriptors like "slender"
  • May vary between images even with same descriptions
• Unwanted tattoos showing up
  • "Skin is clear" and similar terms

General pitfalls

• Images are too closeup/zoomed in
  • You likely aren't describing the whole scenery in some descriptive form. Saying "zoomed out, wide angle" may not work reliably. Instead, describe what they're wearing, especially things like pants & shoes and other background peripherals. The AI will paint what exists in the prompt so if those parts exist then it'll fill in the rest of the image with a fuller scene than closeup.
  • This also applies to blurry backgrounds (bokeh). While natural in photography, if you want high depth of field/low background blur, describe what should be in the background. This will make what you describe clearer and cause less blur.

NSFW Prompts & NSFW engine

• You can toggle using the NSFW engine for specific prompts in the "additional controls" section of single & group selfies.
• NSFW prompts may have more mutations in group selfies - they are better in single person. For groups, best to use a pose reference (and singles too).
• NSFW autoselfies, especially couple autoselfies, are more prone to mutations, and you have an option in autoselfie advanced settings (within the autoselfie menu) to turn on/off NSFW engine. See more in the text in app.

Creative Art Tips

• Weight styles heavily and use style references for abstract art
• Reinforce style by describing it at both start and end of prompt