Atelier Selfies Guide
What’s Different in Atelier#
Atelier is our most advanced engine for identity stability and realism, prioritizing Source-to-Selfie consistency. It uses the entire avatar image which includes skin texture, hair flow, and precise features to define the Kindroid
Atelier follows a strict top-down interpretation of data. Information provided earlier in your setup carries more weight than information provided later. The system follows this hierarchy:
- Avatar Photo (AP): Defines the core identity, including facial structure and hair.
- Avatar Description (AD): Reinforces non-facial traits like body type and skin tone.
- Selfie Prompt: Directs the specific scene, action, and lighting.
The Avatar Photo (Most Important Step)#
Because Atelier maps faces directly from the image, your AP is the single most important factor for success.
- Straight Frontal View: Straight-on views map most accurately. Angled faces, tilted chins, or views from above/below often cause "face drift" or proportionality issues like "bobbleheads".
- High Resolution: The higher the resolution, the better the image detail you can achieve.
- Natural Texture: Avoid overly filtered or smooth AI photos. Visible pores, subtle lines, and natural lighting improve realism and combat “plastic” skin.
- Minimal Background: Use solid colors with high contrast. Cluttered backgrounds can confuse the engine during facial mapping.
- Shoulder Width: Ensure the AP shows a natural shoulder width. Because the engine maps a 2D model, narrow or angled shoulders in a photo can lead to the head appearing disproportionately large in selfies.
Here’s a great prompt to run in Tableau to start with:
A close-up portrait against a simple, muted background. The lighting is bright but diffused, highlighting skin texture. The composition is intimate, focusing on face and shoulders, with a shallow depth of field that subtly blurs the background.
Writing a Precise Avatar Description The Avatar Description (AD) should act as a structural anchor for identity markers that a photo cannot fully convey. Avoid "flowery" or decorative prose.
AD Baseline Template#
- Anchor the Identity: Always explicitly state gender and age to prevent the engine from defaulting to incorrect appearances during complex prompts. “Thirty-two year old man”
- Do Not Describe Facial Structure: Descriptions of eyes, nose, or jawline (with the exception of eye color) are redundant and often create a "collision" where the engine tries to blend two different faces.
Credit Usage#
Single Kindroid = 1 credit
Two Kindroids (or Kindroid + User) = 1.5 credits
Groups of three = 2 credits
Advanced Configuration Tags#
Atelier uses specific syntax to fine-tune facial blending and engine behavior and can be used in the Avatar Description and/or the image prompt.
Head & Face Injection: <...>
Wrap specific attributes in angle brackets to bypass strict facial and expression rendering. This is the required method to control:
- Hair: Use <hair color/style> to drive hair rendering.
- Eyes: Force specific colors or heterochromia (e.g., <left eye blue, right eye green>).
- Expressions: Direct specific gazes or emotions manually.
Face Context Configuration: <(ext:N)> This value controls the "mask" size around the face.
- Default: 1.6 (Range: 1.1 - 3.0)
- Lower (–): Increases facial detail and sharpness but may result in a "pasted-on" look.
- Higher (–): Improves blending for voluminous hair or specific lighting but increases the risk of a "bobblehead" effect. Note: This tag is only read from the AD, not per-shot prompts.
Think of <…> as giving the engine permission to alter that part of the avatar image.
The no-sys Directive Adding no-sys to the AD or prompt removes Kindroid’s own underlying system prompt from the directives. Use this to loosen the engine's "grip" on the AP/AD for more expressive poses, non-human kins, and extreme emotions.
How to Prompt#
Atelier performs best with minimal, structured prompts.
Structure: [Shot Type] of [Identity], [Action/Expression], [Setting], [Clothing], [Lighting/Camera].
- Describe the Whole Scene: If you want a full-body shot, you must describe the legs, shoes, and background to prevent the camera from defaulting to a close-up.
- Lighting Cues: Use "overcast sky" for matte, realistic skin or "cinematic lighting" to avoid flat renders.
- Clothing: If you prefer "passport" style close-ups, avoid mentioning footwear or full-body outfits in the AD, as the engine will try to include them and zoom out.