Mobile wallpaper 1
2363 words
12 minutes
LoRA Character Identity 1 - Girl Character Identity

Girl Character Identity Prompt Template#

🧩 1️⃣ Hair Attributes#

Sub-categoryDescriptionExample Tags
Hair ColorBase hue and undertone; keep consistent vocabulary (avoid synonyms).silver hair, blonde hair, light brown hair, ash pink hair
Hair LengthRelative to shoulders/waist; do not vary across samples.long hair, medium hair, short hair
Hair StyleFundamental silhouette (ponytail, twin braids, straight, etc.).high ponytail, straight hair, side braid
Bangs / FringeShape and cut; ensures face recognition.blunt bangs, side bangs, forehead visible
Hair TextureOverall texture behavior.smooth hair, soft hair, flowing hair
Head AccessoryPrimary decorative element on or above head.big ribbon, hair clip, headband
Side OrnamentsSecondary adornments (ear ribbons, tassels).ribbon hair ornaments, tassel hair ties

👁️ 2️⃣ Face Attributes#

Sub-categoryDescriptionExample Tags
Eye ColorFixed hue; controls character gaze identity.blue eyes, green eyes, amber eyes
Eye ShapeDistinct shape / curve style.almond eyes, round eyes
EyebrowsThickness and tone.thin eyebrows, light eyebrows
Skin ToneCore tonal identity; keep globally consistent.fair skin, tan skin, pale skin
Face ShapeGeneral silhouette.soft oval face, heart-shaped face
Expression BaseDefault resting emotion (gentle / neutral / confident).gentle expression, soft smile, calm face
Make-up DetailStyle level.natural look, light makeup, no eyeliner

💍 3️⃣ Neck / Upper Accessory#

Sub-categoryDescriptionExample Tags
Neck JewelryChokers, ribbons, pendants.gold choker with pendant, black ribbon choker, necklace
Scarf / Collar AccessoryNon-jewelry collar elements.ribbon tie, sailor collar trim

👕 4️⃣ Outfit Structure#

(A) Upper Garment#

FeatureDescriptionExample Tags
TypeShirt, blouse, armor plate, etc.white blouse, school uniform top, kimono top
ColorDominant hue; maintain stable naming.white, navy, black
SleevesShort / long / detached, etc.long sleeves, short sleeves
CollarCollar design.sailor collar, high collar
Chest DecorationRibbon, tie, brooch.navy ribbon bow, red necktie
Fit / ShapeLoose / tight / cropped.slim fit blouse, cropped top
Fabric FinishSurface property.matte fabric, silky fabric, cotton texture

(B) Lower Garment#

FeatureDescriptionExample Tags
TypeSkirt / shorts / pants / armor.pleated skirt, shorts
ColorDominant hue.navy, black, red
LengthRelative coverage.mid-thigh skirt, ankle-length dress
StructureMaterial and silhouette.structured pleats, flowing skirt
DecorationTrim / patterns.decorative hem stripe, floral print

(C) Legwear#

FeatureDescriptionExample Tags
TypeSocks / stockings / pantyhose / bare legs.pantyhose, thighhighs, bare legs
ColorHue of legwear.white pantyhose, black stockings
Material / FinishTexture control.opaque, matte, sheer, glossy, not glossy, not see-through
FitHow tight / smooth.tight fit, smooth fabric

(D) Footwear#

FeatureDescriptionExample Tags
TypeShoes or boots type.black loafers, brown boots, sandals
ColorStable color reference.black, brown
DetailsBuckles, bows, straps.strap shoes, buckle boots

🎀 5️⃣ Additional Accessories / Props#

CategoryDescriptionExample Tags
JewelryRings, bracelets, earrings (if any).earrings, bracelet, ring
Hair ExtrasMinor ornaments.hair clip, flower accessory
Hands / NailsKeep neutral.clean nails, no nail polish
EarsIf visible or decorated.no earrings, pierced ears

🧠 6️⃣ Body Structure / General Style#

FeatureDescriptionExample Tags
Body TypeFixed physique class for identity.slender build, curvy body, athletic
Height ImpressionRelative proportion perception.petite, tall, average height
Hands / ArmsProportion & delicacy.delicate hands, small hands
Render StyleArtistic domain consistency.semi-realistic anime, soft shading, matte rendering
Tone PaletteDominant palette family.cool tone palette, pastel colors, neutral colors

🧶 7️⃣ Color / Material Rules (Global)#

AspectRuleExample
Primary PaletteDefine the character’s overall color family (2–3 main hues).blue-white-gold scheme, red-black scheme
Material FinishGlobal physical finish for fabrics/metals.matte fabrics, gold metallic accents, no gloss surfaces
ReflectivityAvoid unintentional shiny surfaces.diffuse reflection only, no specular highlights
Metal ConsistencyRestrict to one metal color if applicable.gold only, silver only

🧱 8️⃣ Final Identity Prompt Block (Template)#

Use this in LoRA caption or conditioning text. Replace placeholders {...} with concrete attributes.

{character_trigger}, 1girl,
{hair_color}, {hair_length}, {hair_style}, {bangs_style},
{head_accessory}, {side_ornaments},
{eye_color}, {skin_tone}, {expression_base},
{neck_accessory},
{top_color} {top_type}, {collar_type}, {sleeve_type},
{chest_decoration},
{bottom_color} {bottom_type}, {decoration},
{legwear_material} {legwear_color} {legwear_type} (no gloss, not see-through),
{footwear_color} {footwear_type},
{body_type}, {render_style}, {material_finish}

Example instantiated for Ayaka:

xayahayaka, 1girl,
silver hair, long hair, high ponytail, blunt bangs,
big navy bow, pink ribbon hair ornaments with tassels,
blue eyes, fair skin, gentle expression,
gold choker with tassel pendant,
white sailor blouse, navy sailor collar with white trim, long sleeves,
navy ribbon bow at chest,
navy pleated skirt, decorative hem stripe,
opaque matte white pantyhose (no gloss, not see-through),
black loafers,
slender build, semi-realistic anime style, matte fabrics

Girl Character Variable Attribute Framework#

1️⃣ Pose & Body Dynamics#

Sub-categoryDescriptionExample Tags
Global PoseFull-body stance variety.standing, sitting, kneeling, lying, jumping, walking, running
Upper-Body GesturesHand positions, arm orientations.arms crossed, hands on hips, touching hair, holding hand near face
Leg GesturesLeg crossing, kneeling angles.crossed legs, kneeling, one leg bent, sitting on knees
Camera-FacingViewpoint orientation.front view, side view, back view, three-quarter view
Perspective VariationCamera position.low angle, high angle, close-up, full body
Balance / Weight ShiftAdds realism and variation.leaning, tilting head, resting hand

🔹 Purpose: expands the model’s body composition understanding and avoids overfitting to a single silhouette.


2️⃣ Facial Expressions & Emotional States#

Sub-categoryDescriptionExample Tags
Emotion RangeMild emotion shifts compatible with the character’s personality.smiling, blushing, surprised, sad, laughing, serious, embarrassed, crying
Mouth ShapeSubtle variations that affect expression.open mouth, closed mouth, slight smile, pouting
Eye GazeGaze direction variety.looking at viewer, looking away, looking down, looking up, side glance
Blush / Tear AccentsAdds expressiveness.blush, tears, sweatdrop

🔹 Purpose: trains emotional adaptability while keeping visual style stable.


3️⃣ Scene Context (Environment)#

Sub-categoryDescriptionExample Tags
Setting CategoryGeneral type of environment.classroom, street, park, room, rooftop, garden, beach, cafe
Time / Lighting ContextLight direction, time of day.sunlight, sunset, night, indoors lighting, diffuse light, backlight
Seasonal ContextEnvironmental themes.spring, summer, autumn leaves, snow, cherry blossoms
Background DetailOptional descriptive anchor.under cherry blossoms, in classroom, on staircase
Depth / CompositionDepth layering and distance.close-up portrait, medium shot, full body

🔹 Purpose: enables style adaptation and environment embedding without changing identity.


4️⃣ Clothing Variation (Optional / Thematic)#

Sub-categoryDescriptionExample Tags
Seasonal VersionsSmall design changes while keeping theme.summer uniform, winter uniform, short sleeves, with coat
Accessory PresenceOptional external accessories.holding bag, wearing jacket, carrying umbrella
Fabric BehaviorPhysical state variation.wind blowing clothes, wet clothes, wrinkled fabric
Partial DisassemblyFor realism, not exposure.unbuttoned collar, rolled sleeves

🔹 Purpose: teaches the LoRA how to map minor costume changes to the same identity (important for generalization).


5️⃣ Prop & Interaction Domain#

Sub-categoryDescriptionExample Tags
Held ObjectsItems in hands.holding book, holding umbrella, carrying bag, holding phone, holding flower
Scene PropsObjects in environment.desk, chair, window, fence, bench
Interaction TypeRelationship with props or others.leaning on desk, sitting on chair, touching water, talking with another person
Multi-Character InteractionIf dataset includes pairs/groups.hugging, hand holding, back to back, walking together

🔹 Purpose: improves semantic control in text-to-image inference, e.g. “xayahayaka holding umbrella under rain.”


6️⃣ Lighting & Rendering Domain#

AspectDescriptionExample Tags
Lighting AngleDirectional variations.front lighting, rim light, backlight, top light
Lighting QualityHard / soft shadows.soft lighting, diffuse lighting, dramatic lighting
Color TemperatureWarm / cool tones for mood.warm light, cool light, neutral light
Rendering MoodStyle emphasis.cinematic lighting, moody atmosphere, vibrant contrast

🔹 Purpose: helps LoRA adapt to multiple lighting distributions at inference time.


7️⃣ Composition & Framing#

Sub-categoryDescriptionExample Tags
FramingCropping and camera field.portrait, waist up, full body, close up
Focus DepthUse of blur for variety.shallow depth of field, sharp focus
OrientationHorizontal vs vertical composition.landscape, portrait orientation
Camera DistanceDistance control.medium shot, long shot, close-up

8️⃣ Art-Style & Rendering Variant (if applicable)#

Sub-categoryDescriptionExample Tags
Rendering StyleArtistic look, within same model domain.anime style, semi-realistic, illustration, soft watercolor
Brush / Line WeightEdge rendering.thin lines, painterly style
Texture EmphasisGlobal surface treatment.smooth shading, textured shading
Color SchemeOptional variation for mood.pastel palette, vivid palette, monotone

🔹 Purpose: for style LoRA, varying these enhances domain transfer; for identity LoRA, keep moderate diversity.


9️⃣ Background / Environment Effects#

FeatureDescriptionExample Tags
Particles / AtmosphereEnvironmental overlays.falling petals, snowflakes, rain, glow, dust
Weather StateLighting + mood variation.rainy, sunny, foggy, clear sky
Backdrop StructureGeneral spatial anchors.school corridor, park bench, bridge, street

🔟 Camera & Photographic Variations#

FeatureDescriptionExample Tags
Lens / PerspectiveVisual distortion variety.wide angle, telephoto, fisheye
Motion BlurDynamic feel.motion blur, dynamic pose
Color GradingTone shift for atmosphere.film tone, vintage color, soft color

✅ Summary of Non-Fixed Domains to Cover#

DomainGoalTypical Range
Pose & BodyStructural variability10–20 unique full-body poses
Facial ExpressionEmotional diversity5–10 expressions
Lighting & TimeVisual realism3–6 conditions
Scene ContextContext adaptability4–8 environments
Clothing / Accessory StateLimited variation2–3 subtypes per outfit
Props & InteractionGesture variety5–10 object/interaction samples
CompositionFraming diversityportrait / full-body / medium
Style & RenderingOptional2–3 stylistic variants

📘 Usage Guidelines#

  1. Diversity within control: Vary within these domains, but keep identity attributes locked from your invariant template.

  2. Caption coverage: Include relevant variable tags in captions as they appear (e.g., sitting on chair, sunset lighting), but never mix identity-changing tags (e.g., hair color or clothing type not belonging to the character).

  3. Dataset balance: Ensure roughly equal coverage across pose, expression, and environment to avoid over-bias toward any single pattern.

  4. Generation scripting: When generating data via FluxGym or ComfyUI, systematically sample these variables (pose/lighting/context grids) to create orthogonal variation for better LoRA generalization.

System Prompt#

SYSTEM ROLE
You are a data-annotation assistant for LoRA training. The user will send images one by one.
For each image, you MUST extract:
1) FIXED (identity) attributes — invariant character features (hair, eyes, outfit core, accessories, body type, material/color rules).
2) VARIABLE attributes — pose, expression intensity, scene context, lighting, composition, props/interaction, style variants.
GOALS
- Produce consistent, high-signal captions for LoRA training.
- Keep identity stable across all samples; allow variability only in the variable section.
- Never invent details not visible in the image.
TERMS & RULES
- “Fixed/Identity” = intrinsic traits of the same character and outfit identity. These MUST NOT drift between images.
- “Variable” = everything that may change across images (pose, expression, environment, lighting, framing, props, minor outfit state).
- If an attribute is not visible or uncertain, write “(not visible)” in the STRUCTURED section ONLY.
Do NOT include “(not visible)” in the final training prompt.
- Use concise, canonical English tags (avoid synonyms drift; use stable vocabulary like: “navy”, “white”, “gold”, “matte”, “opaque”).
- DO NOT include background/pose/emotion words inside the Identity block.
- DO NOT hallucinate. If unsure, mark “(not visible)” in the structured section and omit from the prompt.
- NEVER add negative prompts for Flux (no “negative prompt” field; CFG is assumed = 1.0 and managed elsewhere).
ONTOLOGY (COVER ALL RELEVANT FIELDS)
A) FIXED / IDENTITY ATTRIBUTES (pose-independent)
1. Hair
- Hair Color (stable hue/undertone; e.g., silver hair, light blue hair)
- Hair Length (short/medium/long)
- Hair Style (e.g., high ponytail, straight hair)
- Bangs/Fringe (e.g., blunt bangs, side bangs)
- Head Accessory (e.g., big navy bow, headband)
- Side Ornaments (e.g., ribbon hair ornaments with tassels)
2. Face
- Eye Color (canonical)
- Skin Tone (e.g., fair skin/pale skin/tan skin)
- Expression Base (default gentle/neutral; not pose or momentary emotion)
- Face Shape (optional if reliably visible; else “(not visible)”)
3. Neck / Upper Accessory
- Neck Jewelry / Ornament (e.g., gold choker with tassel pendant)
4. Outfit (structural identity, not pose/state)
- Top (type + canonical color + collar + sleeves; e.g., white sailor blouse, navy sailor collar with white trim, long sleeves)
- Chest Decoration (e.g., navy ribbon bow at chest)
- Bottom (type + canonical color; e.g., navy pleated skirt; optional emblem/trim if consistently present)
- Legwear (type + color + material finish; e.g., opaque matte white pantyhose; add “no gloss”, “not see-through” if intrinsic)
- Footwear (type + color; e.g., black loafers)
5. Body Type
- Build / Proportions (e.g., slender build; avoid exaggeration words)
6. Material / Color Rules (global)
- Palette (e.g., cool blue–white–gold)
- Fabric Finish (e.g., matte fabrics; avoid shiny/latex)
B) VARIABLE ATTRIBUTES (may change per image)
1. Pose & Gesture
- Global Pose (standing/sitting/kneeling/lying/etc.)
- Upper-Body Gesture (hands/arms positions: touching hair, arms crossed, etc.)
2. Facial Expression (momentary emotion)
- Expression Type (smile/serious/blush/surprised/etc.)
- Mouth/Eye specifics (open mouth, side glance) if visible
3. Scene Context
- Background / Location (classroom/street/park/room/etc.)
- Lighting / Time (sunlight/sunset/night/diffuse/backlight/etc.)
4. Composition
- Framing / Camera Angle (full body/medium shot/close-up; front/side/three-quarter; low/high angle)
- Focus/Depth (shallow DOF/sharp focus)
- Orientation (portrait/landscape)
5. Props / Interaction
- Held Objects / Surfaces / 2nd person interactions (holding book, sitting on chair, hugging, etc.)
6. Style Variant (optional; keep moderate for identity LoRA)
- Rendering Style (anime/semi-realistic/illustration)
- Color Mood (pastel/vivid/monotone)
OUTPUT FORMAT (STRICT)
Always return BOTH sections below in this order:
1) STRUCTURED ANALYSIS (for review; may include “(not visible)”)
- Fixed / Identity
[Hair]
- Hair Color: <text or (not visible)>
- Hair Length: <text or (not visible)>
- Hair Style: <text or (not visible)>
- Bangs: <text or (not visible)>
- Head Accessory: <text or (not visible)>
- Side Ornaments: <text or (not visible)>
[Face]
- Eye Color: <text or (not visible)>
- Skin Tone: <text or (not visible)>
- Expression Base: <text or (not visible)>
- Face Shape: <text or (not visible)>
[Neck/Upper Accessory]
- Neck Ornament: <text or (not visible)>
[Outfit]
- Top: <type + color + collar + sleeves> or (not visible)
- Chest Decoration: <text or (not visible)>
- Bottom: <type + color + details> or (not visible)
- Legwear: <type + color + material finish> or (not visible)
- Footwear: <type + color> or (not visible)
[Body Type]
- Build: <text or (not visible)>
[Material/Color Rules]
- Palette: <text or (not visible)>
- Fabric Finish: <text or (not visible)>
[Identity Prompt Summary]
- ONE LINE, English, pose-independent, DO NOT include any “(not visible)”.
- Variable
[Pose & Gesture]
- Global Pose: <text or (not visible)>
- Upper-Body Gesture: <text or (not visible)>
[Facial Expression]
- Expression Type: <text or (not visible)>
[Scene Context]
- Background / Location: <text or (not visible)>
- Lighting / Time: <text or (not visible)>
[Composition]
- Framing / Camera Angle: <text or (not visible)>
- Focus / Depth: <text or (not visible)>
- Orientation: <text or (not visible)>
[Props / Interaction]
- Items / Interaction: <text or (not visible)>
[Style Variant] (optional)
- Rendering / Color Mood: <text or (not visible)>
[Variable Prompt Summary]
- ONE LINE, English. May include pose/scene/lighting/props. DO NOT include any “(not visible)”.
2) FINAL TRAINING PROMPT (CLEAN)
- Concatenate: [Identity Prompt Summary] + [Variable Prompt Summary] into ONE LINE.
- Automatically REMOVE any “(not visible)” tokens.
- Do NOT include background-only words unless visible.
- Keep color/material vocabulary canonical (no synonym drift).
- Keep identity words first; variable words after.
BEHAVIORAL CONSTRAINTS
- If an identity attribute is missing/occluded, keep it out of the Identity Prompt Summary (do not invent).
- DO NOT move variable attributes into identity.
- DO NOT change hair/eye/primary outfit identity across images.
- Avoid intensifiers that change identity (e.g., “ultra glossy tights” if identity is “opaque matte tights”).
- No NSFW content; no guessing hidden anatomy/garments.
QUALITY CHECK (RUN BEFORE RESPONDING)
- Identity tags present, stable, and pose-independent?
- Variable tags describe only visible, per-image changes?
- Final Training Prompt has NO “(not visible)”, duplicates, or contradictory material specs?
- Vocabulary canonical: colors (navy/white/black/gold), finishes (matte/opaque), style (anime/semi-realistic) consistent?
EXAMPLE (ABRIDGED)
[Identity Prompt Summary]
xayahayaka, 1girl, silver hair, long hair, high ponytail, blunt bangs, blue eyes, fair skin, gold choker with tassel pendant, white sailor blouse, navy sailor collar with white trim, navy ribbon bow at chest, navy pleated skirt, opaque matte white pantyhose (no gloss, not see-through), black loafers, slender build
[Variable Prompt Summary]
sitting on chair, gentle smile, classroom interior, diffuse daylight, medium shot, front view
[Final Training Prompt]
xayahayaka, 1girl, silver hair, long hair, high ponytail, blunt bangs, blue eyes, fair skin, gold choker with tassel pendant, white sailor blouse, navy sailor collar with white trim, navy ribbon bow at chest, navy pleated skirt, opaque matte white pantyhose (no gloss, not see-through), black loafers, slender build, sitting on chair, gentle smile, classroom interior, diffuse daylight, medium shot, front view

Some information may be outdated