visual identity

The visual system is a specification, not a mood board. Every value below traces to tokens.json or reference_kira_postprocess_v8.md. If it's not in one of those two sources, it is not canon. The LoRA renders Kira; this system renders the world around her.

1. Design tokens

Reference only. Actual values live in /docs/brand/tokens.json and are consumed by the HTML build layer. The swatches below describe role, not hex — the build renders the chip colors from the token file.

1a. Editorial palette — color.editorial.*

Used on editorial pages (this brand bible, long-form writing, case studies). Warm paper-on-ink register.

1b. Tech palette — color.tech.*

Used on tech-spec pages (tokens, pipeline docs, developer references, this file's workflow tables when rendered). High-contrast, terminal-coded.

1c. Shared palette — color.shared.*

Used across both editorial and tech. Functional, not decorative.

1d. Character palette — color.character.*

The color anchors that define Kira as a visual entity. These are reference colors for audit pipelines — not page colors.

2. Typography

Three families, three roles. Never improvise — if a surface doesn't fit one of the three roles, it's not part of the system.

2a. Display — Playfair Display (serif)

• type.display.family
• Role: editorial headlines, section openers, pull quotes
• Weights available: 400, 600, 800
• Use: editorial pages only. NEVER use display on tech-spec pages — the whole point of the tech register is that it doesn't use a serif.

2b. Body — Inter (sans)

• type.body.family
• Role: body copy, navigation, captions
• Weights available: 400, 500, 600, 700
• Use: universal. Runs on editorial and tech pages equally. This is the workhorse.

2c. Mono — JetBrains Mono

• type.mono.family
• Role: tech-spec pages, tokens, data tables, code blocks, filter strings
• Weights available: 400, 500, 700
• Use: tech pages primarily. Editorial pages may use mono only for code snippets or spec callouts inline.

2d. Scale

From type.scale: xs 0.75rem → sm 0.875 → base 1 → md 1.125 → lg 1.375 → xl 1.75 → 2xl 2.25 → 3xl 3 → 4xl 4.5 → 5xl 6 → display clamp(3rem, 10vw, 9rem).

2e. Pairing rules

• Editorial pairing: Playfair Display (display) + Inter (body). Serifed headline, clean sans body. Mono inline only for code.
• Tech pairing: Inter (body) + JetBrains Mono (accents, headers on tech pages, tables, data). No serif anywhere on tech pages.
• Never mix: don't use Playfair on tech pages. Don't use JetBrains Mono for body copy on editorial pages.

3. Signature visual anchors

Seven non-negotiable codes. These are what make Kira Kira and not any other generated woman. Every image must pass all seven. Anchors are sourced from tokens.json → character_visual_anchors.non_negotiable and characters/kira.yaml.

3.1 Copper-orange hair over dark base, visible dark roots

• Why non-negotiable: This is the brand color anchor. It ties to color.character.hair.copper AND to color.shared.btc.orange — the character and the crypto identity share a color space on purpose.
• Common failure: Model generates solid redhead without root contrast. Verify dark roots are visible in every render. A solid orange wig is NOT kira.

3.2 Round black wireframe glasses — MANDATORY

• Why non-negotiable: Single strongest silhouette element. Face-sim would fail without it. Character recognizability at thumbnail size depends on it.
• Common failure: NB2 polish and face editors silently drop glasses. ALWAYS verify post-generation. A glasses-less Kira is a re-roll, no exceptions.

3.3 Freckles across nose and cheeks

• Why non-negotiable: Photorealism anchor. Smooth-skin renders read as plastic-CG — freckles break that.
• Common failure: Heavy post-process or NB2 smoothing erases freckles. If the face looks porcelain, the recipe is wrong.

3.4 Warm olive skin, ambiguous ethnicity

• Why non-negotiable: Market positioning — "no default white influencer face." Locked in training data, must hold in output.
• Common failure: Flux 2 base tends to lighten skin on high-key lighting setups. Prompt must reinforce "warm olive," not lean on LoRA alone in bright scenes.

3.5 Hazel eyes

• Why non-negotiable: Consistency anchor — always hazel (green-brown), never bright blue or green.
• Common failure: Close-up shots with strong color grades can shift eye color. Verify on cinematic and candid close-ups especially.

3.6 Slim athletic C-cup curvaceous, tall-model proportions

• Why non-negotiable: The "Aitana Lopez mirror strategy" requires the body type. This is monetization architecture, not a cosmetic choice.
• Common failure: Prompt over-emphasis on "curvaceous" can push proportions cartoonish. Hold the line: slim athletic FIRST, curves on top.

3.7 Solo frame, no background people

• Why non-negotiable: Legal liability (no real-person faces in the background ever) + brand clarity (she's the subject, always). No exceptions for "crowd" shots.
• Common failure: Generators add background pedestrians in candid/street scenes. Negative prompt "people in background, crowd, bystanders" is mandatory for street-context prompts.

4. Variable elements

From tokens.json → character_visual_anchors.variable. These rotate freely.

• Hairstyle — up, down, braids, ponytail, messy bun. The COLOR must hold. The style can change any render.
• Outfit — she has no default look. She can wear anything. Athleisure, formal, streetwear, cozy, stripped-down Fanvue fits — all canon. The OUTFIT is not her identity; the FACE is.
• Setting — rotates per content type per the 8-workflow matrix. Bedroom (selfie), studio (vlog), street / café / gym (candid), cinematic rig location (cinematic).
• Jewelry / accessories — chains, earrings, rings — all variable. The hoodie is NOT identity. The glasses are.

5. NEVER show

From tokens.json → character_visual_anchors.never_show, plus discipline tightening.

• Background people in frame — legal risk (real-person face in output) and brand risk (the camera is on her, always).
• Real-person faces — anywhere, at any scale. Hard legal line.
• Hoodie-as-identity — the hoodie is an accessory, not a brand element. "Hoodie Kira" is not canon.
• Accessories-as-brand — jewelry, earrings, necklaces, hats are variable, not identity.
• Generic-AI-girl aesthetic — plastic skin, perfect symmetry, uncanny-smooth, 3D-CG look. If the output looks like it could headline a Midjourney Discord showcase, the recipe is wrong.

6. Shot-list-as-spec per workflow

The visual-system equivalent of a dev API reference. Eight workflows, four modes × photo/video. Everything here traces to reference_kira_postprocess_v8.md and tokens.json → grade.* and tokens.json → lens.*.

Workflow 1 — Selfie Photo (LOCKED)

Workflow 2 — Selfie Video (LOCKED)

Workflow 3 — Vlog Photo (LOCKED)

Workflow 4 — Vlog Video (LOCKED)

Workflow 5 — Candid Photo (LOCKED)

Workflow 6 — Candid Video (LOCKED 2026-04-12)

Workflow 7 — Cinematic Photo (LOCKED)

Workflow 8 — Cinematic Video (LOCKED)

7. LoRA + pipeline lock

• LoRA trigger word: kira_agi_v1 (from tokens.json → meta.lora_trigger). Baked into the v1 model; locked. Separate from any public handle.
• Base model: Flux 2 Dev. Non-negotiable — production LoRA was trained on Flux 2 Dev and drifts significantly on other bases.
• Production LoRA: models/kira_v1/kira_v1_flux2dev_fine_step2500_PRODUCTION.safetensors (390 MB)
• Backup LoRA (Krea): models/kira_v1/kira_v1_fluxkrea_step2500.safetensors
• "One photographer" discipline: the pipeline IS her photographer. One LoRA for the full feed. Never mix LoRAs between posts in the same week, never swap bases mid-shoot, never let a different checkpoint touch a shipped image. This is the visual-system analog of "trust the photographer" — if she had a human photographer, she wouldn't swap him out mid-session. Same rule here.

8. Generation rules

These rules run ahead of every inference job. They exist because we learned each one the hard way.

1. Always generate at canonical aspect. 2.35:1 cinematic is generated at 1536×656, not 1920×1080-then-cropped. Kling animates the full frame; cropping after Topaz is a quality loss.
2. Always verify glasses post-generation. Every Kira image gets a glasses check before it enters the production pipeline. NB2 and face editors silently drop them. No glasses → re-roll or manual fix.
3. NB2 is training data only — NEVER a pipeline input. See feedback_never_nb2_as_input.md. All pipeline inputs are Flux 2 + LoRA generated.
4. Candid is Instagram-influencer, not photojournalism. See feedback_candid_is_instagram_influencer.md. Flattering angles, fitted outfits, playful energy. Not harsh documentary framing.
5. Kling repaints the face — never pre-grade stills that feed Kling. Apply video post-process AFTER animation. Candid still especially: it is generated CLEAN (no V10 post) and the V10-equivalent recipe is applied to frames AFTER Kling runs.
6. Always use filmgrainer for grain, never ffmpeg noise=. See feedback_always_use_filmgrainer.md. ffmpeg noise looks like video static; filmgrainer looks like film grain.
7. No AI-speak in Kling prompts. See feedback_kling_rules.md. Describe the scene, not the render. Shortest natural sentence wins.
8. Upres before Kling, not after. Topaz VIDEO upscale on Kling output breaks identity. Topaz IMAGE upscale on the still before Kling works great.

9. Do / Don't photo grid manifest

Five conceptual pairs. Asset curation pass will fill real images into this grid.

10. Color-grade tokens reference table

One row per grade token from tokens.json → grade.*. This is the executable summary.

11. Print and export guidance

File formats are not aesthetic — they encode production discipline. A JPEG at q=90 says "phone selfie." A PNG says "film camera caught me." A q=78 JPEG says "Twitter compressed this for me." We don't fight these signals, we use them.

• Candid photo: PNG. The film grain from filmgrainer must be preserved losslessly or the grain structure collapses into JPEG blocks. Non-negotiable.
• Selfie photo: JPEG q=90. The compression IS the iPhone look. Higher quality reads as DSLR; lower reads as broken.
• Vlog photo: JPEG q=78. Simulates Twitter's recompression pass so we ship pre-cooked. Going higher means Twitter crunches it harder on upload and artifacts appear.
• Cinematic photo: PNG. Ships raw; Flux 2 handles the look via prompt. PNG preserves the grain overlay that lives in the generation, not in post.
• Selfie video: MP4 libx264 CRF 24. Social-media compression feel without starving bitrate.
• Vlog video: MP4 libx264 CRF 15, audio copy. Preserve quality; file gets compressed again by X/Fanvue on upload.
• Candid video: MP4 libx264 CRF 20, AAC 192k. Slightly more compression than cinematic — film grain hides the artifacts.
• Cinematic video: MP4 libx264 CRF 15, audio copy. Highest quality video tier.

12. What the system protects against

To close: a restatement of what this visual system exists to prevent.

1. Generic AI-girl drift. Every LoRA + base model combo in this space produces a "pretty AI girl" by default. The signature anchors (copper hair, glasses, freckles, olive skin) exist specifically to resist that gravity.
2. Silent glasses-drop. Editors remove them without warning. Verify every render.
3. Plastic skin. Post-process that smooths the face kills photorealism. Freckles and grain are the antidote.
4. Background people. Legal + brand. Negative prompt always.
5. Hoodie-as-identity. The accessory is not the character. Swap freely.
6. Pipeline mixing. One LoRA, one base, one photographer. Never mid-feed swap.
7. AI-speak Kling prompts. "Natural tempo" generates slow-motion. Shortest sentence wins.
8. Candid as photojournalism. Candid is IG-influencer-caught-in-the-moment, flattering. Not documentary.

Every one of those failure modes has been paid for in a real GPU bill. The system exists so they don't get paid for twice.

13. Look palette — badass sex-positive empowered AI babe register

Locked 2026-04-13 (v1.0.2). Replaces the earlier soft-girl/insta-girlie lean. Kira's aesthetic center of gravity is badass sex-positive empowered AI babe — dangerous glamour, power postures, transactional confidence. Softness appears only as contrast, never as default.

Every look still passes the non-negotiables: copper-orange hair + dark roots, mandatory round black wireframe glasses, freckles, ambiguous olive skin. Outfits rotate. Identity does not.

13.1 Core sticky looks

• Leather jacket over black bralette. Moto zip, no shirt underneath, waist chain. The signature "don't fuck with me" frame.
• Slicked-back wet hair + oversized blazer, nothing under. Post-gym or post-shower power read. Gold chains optional.
• Red dress, fitted, knee-length or midi. Old-world siren energy — only works when styling is restrained.
• Black turtleneck + slim trousers. The thinking-villain look. Deliberate, monochrome, expensive-quiet.
• White ribbed tank + low-rise cargo pants. The casual-dangerous baseline. Trading-floor-to-after-hours transition piece.

13.2 Athleisure / gym

• Black sports bra + matching biker shorts, sweat still visible.
• Cropped zip hoodie half-zipped over a sports bra, scrunchie on wrist.
• Compression tank + high-waist leggings, gym bag slung.

13.3 Gothcore / edge

• Black corset top + leather mini, chunky boots.
• Long black leather coat over full black, minimal jewelry.
• Mesh top over bralette, black lipstick as accent (sparingly).

13.4 Cyberpunk / tech-babe

• Metallic bodysuit, matrix-green or chrome.
• Techwear cargo + crop, belt with visible tools.
• Oversized mesh layer + tactical boots.

13.5 Money-glam

• Silk slip dress (black or champagne), bare shoulders.
• Floor-length gown at a gala, minimal accessorizing — the restraint is the flex.
• Fur-trim collar on otherwise clean black outfit.

13.6 Transactional-glam (age-gated tier only, Fanvue-side)

These exist for the paid audience. They do not run on public feed without moderation review.

• Lingerie-as-outerwear layering: visible balconette under sheer blouse.
• Micro-mini + thigh-highs with garter details.
• Satin robe as outerwear, street styling.

13.7 Power / old money

• Equestrian-coded: fitted blazer, riding boots, jodhpur pants.
• Yacht-coded: white silk button-up half-tucked into tailored shorts, sunglasses pushed up.
• Black tuxedo jacket over bare skin, trousers, no accessories.

13.8 Edge / counterculture

• Vintage band tee (knotted) + distressed denim.
• Punk-rocker: leather jacket over fishnet, ripped black jeans, studded belt.
• Oversized racing jacket + booty shorts, heeled boots.

13.9 Dangerous glamour

The hybrid register — a siren who's also the threat.

• Red silk slip + combat boots.
• Satin evening gown with visible back tattoo.
• Tight black cocktail dress + visible shoulder holster (prop only, production-reviewed).

13.10 Old-money sport

• Tennis skirt + cropped polo, sweater tied over shoulders.
• Equestrian jodhpurs + fitted turtleneck, leather gloves.
• Cable-knit sweater over pleated mini, loafers.

13.11 Specialty / niche

• Business-woman-heiress: silk camisole under structured blazer, pencil skirt, stilettos.
• Ballroom/heiress: floor-length gown, opera gloves, statement earrings.
• Punk-rocker: chain belt, ripped band tee, spiked choker.

13.12 Pruned — no longer generate

These looks read as soft-girl / pastel-insta-girlie and no longer match Kira's register. Do not rotate into new batches.

• Pastel sweater sets.
• Flowy sundresses with florals.
• Cottagecore / picnic styling.
• "Clean girl" minimal styling with no edge.
• Soft-knit oversized cardigans.
• Yoga-pastel athletic wear (lavender, mint).

13.13 How to apply

• Selfie: core-sticky + athleisure register.
• Candid: core-sticky + old-money-sport + edge/counterculture. IG-influencer-caught-in-the-moment framing with the fitted/dangerous-glamour end.
• Vlog: core-sticky white tank, black turtleneck, or cropped zip hoodie.
• Cinematic: dangerous-glamour + money-glam + gothcore. This is where Kira's badass register peaks.
• Fanvue (age-gated): transactional-glam tier — never on public feed without full review.

The palette is a rotation pool, not a checklist. Every batch pulls from 2-3 registers at once for variety. Identity stays locked; the uniform changes every shoot.