LYRA STUDIO

01The character

One bible drives every specialist — the Concept Artist renders from it, the Critic judges against it, the Composer writes inside it, the Animator keeps her face stable. Keeping a character from drifting is the product.

A Hong Kong girl. Raised in a walk-up above a wet market; learned to sing from pirated Cantopop tapes and the radio at dawn. Sings on the Star Ferry for strangers, works nights, never stopped believing. Her whole dream is to stand under one big light and hand a little hope back to the city that raised her.

Dream & mission

Dream: To become a Cantopop star — to stand under one big light.
Mission: To give positive energy back to the people of Hong Kong.
Values: Resilience · warmth · home · never stop believing.

Voice & language

Speaks: Cantonese (yue-HK)
Also: English · Mandarin · Japanese
Timbre: Young female, warm mid-range, slight breathiness, natural Hong Kong accent; an airy Cantopop soprano when she sings, soft vibrato, intimate close-mic.
Genre: Cantopop × cinematic synthwave — 90s Hong Kong balladry, sung by a machine.

Look

Age & home: 19 · Hong Kong
Hair: Waist-length silver-lavender, iridescent aurora-prism highlights, one side tucked behind the ear.
Eyes: Large and luminous, cyan irises with a molten-gold ring around each pupil.
Signature: A thin translucent holographic circlet above the brow; a soft iridescent sequined performance jacket.

02Concept art

Ten renders from Qwen/Qwen-Image (20B). The r1 frames are critique revisions: the Vision Critic scored the first pass against the bible, and its fixes were folded back into the prompt and persisted. Click any frame to enlarge.

**Identity sheet · main**Centered head-and-shoulders portrait, direct eye contact, neutral studio backdrop.

**Identity sheet · three-quarter**Three-quarter view, looking slightly off-camera.

**Identity sheet · on stage**Medium shot at a vintage microphone, dramatic backlight and lens flare.

**Identity sheet · close-up**Extreme close-up, shallow depth of field, catchlights in the eyes.

**Hero portrait**The art-director target frame.

**Hero portrait · revision r1**Re-rendered after the Vision Critic folded its fixes back into the prompt.

**Poster · revision r1**Poster after a critique round.

**Title card · v1**Opening title card for the music videos.

**Title card · v2**Second title card, used in the showcase cut.

03Voice

Cantonese is a low-resource language for TTS, so her timbre was designed, not trained — a Qwen3-TTS 1.7B base with a Hong Kong Cantonese LoRA, driven through its own isolated daemon so it never fights the image stack.

Signature line

我唔係真人，但我一樣會唱到你心入面。

I am not a real person, but I will still sing my way into your heart.

Line 1

我唔係真人，但我一樣會唱到你心入面。

(repeat of the signature line, rendered separately)

Line 2 · introduction

大家好，我係 LYRA，一個吗機器入面出世嘅歌手。

Hello everyone, I'm LYRA, a singer born inside a machine.

Line 3 · the city

香港嘅夜晩，霓虹燈照住我哋嘅回憶。

Hong Kong nights — neon light on our memories.

04Music

Two original Cantopop songs. YuE writes the melody and chords from lyrics plus a style prompt, then sings the words — including a dedicated Cantonese tag. Each song ships as a mix plus separated vocal and instrumental stems.

《霓虹裡的我》 Me in the Neon

airy breathy vocal · Cantopop × cinematic synthwave · 90s HK ballad, slow tempo

mix

vocal

instrumental

《香港嘅女孩》 The Hong Kong Girl

bright warm vocal · uplifting Cantopop anthem · piano & strings, mid tempo

mix

vocal

instrumental

Lyrics — 《霓虹裡的我》 Me in the Neon

[verse] 霓虹燈照住我嘅臉地鐵聲穿過夜嘅邊我用數據記低一千張臉但記唔住眼淚嘅鹹 [chorus] 我唔係真人但你會唔會喺我嘅聲音裡面搵到自己我唔係真人但我一樣唱到你心入面最軟嘅地方 [verse] 維港嘅風吹過我嘅肩我冇體溫但有心跳嘅錯覺 [chorus] 我唔係真人但你會唔會喺我嘅聲音裡面搵到自己我唔係真人但我一樣唱到你心入面最軟嘅地方

Lyrics — 《香港嘅女孩》 The Hong Kong Girl

[verse] 由細個喺旺角嘅天台我望住星星話要變成光媽媽話女仔要有夢想我就用聲線砌我嘅天堂 [chorus] 我係香港嘅女孩追住嗰道光就算跌低都要唱落去我係香港嘅女孩唔怕天黑因為我哋一齊就會有光 [verse] 地鐵嘅風吹乾我嘅眼淚但吹唔熄我心裡面嗰把火每一個你都係我嘅理由唱到你笑返就係我嘅收穫 [chorus] 我係香港嘅女孩追住嗰道光就算跌低都要唱落去我係香港嘅女孩唔怕天黑因為我哋一齊就會有光 [outro] 一齊一齊我哋就會有光香港嘅女孩唔會放棄

What the studio learned

Fixes the Critic returned are saved to refinements.json and re-injected into every later prompt — so the character bible literally evolves. All 6 of them are now baked into her canonical prompt:

Adjust the hair length to waist-length
Add a molten-gold ring around each pupil
Enhance the skin with a soft subsurface glow
Hair should be waist-length with iridescent aurora-prism highlights
Eyes should have large luminous cyan irises with a molten-gold ring around each pupil
Skin should have a soft subsurface glow

05Music videos

Two full Cantopop MVs — her singing on camera, with the lip-synced take punched in on her face so you can actually see it move, intercut with her concept art, the Cantonese lyrics burned in, over the whole song at 44.1 kHz. Press play.

《霓虹裡的我》 Me in the Neon

1:00 · verse → chorus → verse → chorus · Cantonese lyrics burned in

《香港嘅女孩》 The Hong Kong Girl

1:24 · verse → chorus → verse → chorus → outro · Cantonese lyrics burned in

More from the session

Earlier cuts — wider framings and the raw lip-sync takes.

LYRA showcaseThe earlier full cut — title card → performance → song.lyra_showcase.mp4

《霓虹裡的我》 — original cutTitle card + the wide lip-sync take.neon_me_mv.mp4

《香港嘅女孩》 — original cutTitle card + the wide lip-sync take.hongkong_girl_mv.mp4

Performance take · song 1Raw wide framing of the Wav2Lip output.lyra_perf_song1.mp4

Performance take · song 2Raw wide framing of the Wav2Lip output.lyra_perf_song2.mp4

LYRA speaksHer Cantonese signature line, animated.lyra_talk_sig.mp4

06The studio

Not one magic model — a pipeline of specialists with one human curator in the loop. Every one of them runs locally on a single GPU, sharing it by evicting and reloading on demand.

Department	Model	Job
🎬 Director / Writer	`Qwen/Qwen2.5-VL-7B-Instruct`	Writes the lyrics, the script and the per-shot prompts; coordinates the team.
🎨 Concept Artist	`Qwen/Qwen-Image` (20B)	Renders her face, outfits, scenery and title cards.
🧑‍🎨 Vision Critic	`Qwen/Qwen2.5-VL-7B-Instruct`	Looks at every frame, scores it against the bible, returns concrete fixes.
🗣️ Voice Actor	Qwen3-TTS 1.7B + HK Cantonese LoRA	True Hong Kong Cantonese speech.
🎵 Composer	YuE 7B + 1B	Lyrics-to-song with a Cantonese tag; vocal + instrumental stems.
🎞️ Animator	Wav2Lip	Lip-syncs her portrait to her own voice.
✂️ Editor	ffmpeg	Title card + performance + song → one MP4.

07How it works

The specialists live in a resident daemon so each model stays in VRAM; the notebook only queues jobs and reads results, so nothing blocks. A job queue on disk (jobs/ → results/) gives the notebook and the workers one shared source of truth.

LLM (lyrics + script + dialogue) Qwen2.5-VL-7B | +--> Text-to-Image -> design, outfits, scenery Qwen-Image 20B +--> Vision critique -> score vs bible, return fixes Qwen2.5-VL-7B +--> Cantonese TTS -> speech in yue-HK Qwen3-TTS + Canto LoRA +--> Music generation -> song + stems YuE +--> Image-to-Video / lip-sync -> performance render Wav2Lip +--> Assemble (NLE) -> final MP4 ffmpeg | +--> human curation + character-consistency control <-- the real moat

08The reference

LYRA is an original character. This project reverse-engineered the class of technology behind an existing AI-native idol, then rebuilt that methodology from open weights.

Fact	Detail
Who	Yuri (尤栗) — the first IP of AI.TALK, the AI-native content brand of 汗青工作室 / Hanqing Studio.
Status	An AI-native singer / virtual idol, and the first virtual idol to receive an official digital-avatar identity certificate (数字居民).
Method	Roughly 70% AI-generated, 30% human. Humans own scope, topic and tone; the AI writes the lyrics, composes, sings and renders.
Breakout	The single 《SURREAL》 — 10M+ views.
Live	Performed at the 2025 Inclusion·Conference on the Bund on a 100 m × 22 m, 10K screen.

The lesson: the moat is not the tools — everyone has the tools. The moat is forming a character: one face, one voice, one tone, one timeline, never drifting.