这段录自网页里实时运行的澄夏，没有一帧是手绘的动画。
像一张会呼吸的
空间照片。
她在浏览器里一帧帧实时画出来，不是预先录好的动画。
头慢慢漂，前后层以不同的速度错开，平面原画有了纵深。
双马尾的聊天里跑的也是这同一套视差代码。
原画加一句话，
就是她的全部行李。
温暖活泼、好奇心很强的旅行者，喜欢听你讲今天发生的事。
图二 她骨架里的眼睛：换表情时逐帧切换
一张设定图里，已经有她的四种样子。
— from the design sheet
双马尾和澄夏走的是同一条流水线、同一套代码。下面是两个角色的全身待机，录自网页里实时运行的画面。
两边的数字不一样，说明骨架不是照抄的：同一套代码读她自己的原画，长出属于她的骨骼和网格。
图三 双马尾骨架里的眼睛
做一个能动的二次元角色，传统上要原画、拆件、补画、绑骨、调动作好几道人工。我们把每一道都交给程序和模型。
身体与斗篷
前臂和手
脸和五官
前发与辫子
左：网页里实时运行的她。上：骨架里的眼睛部件。
每个数字都直接从骨架文件里读出来，不估、不凑。
一步一个脚印，
每一步都有记录。
澄夏诞生：深蓝长发、琥珀色眼睛、旅行斗篷和指南针发饰。
项目第一次入库：聊天网页、脚本和文档一起出发。
澄夏和双马尾第一次用流水线生成的骨架活动，脸是活的。
每条线的成果都放进同一个可交互的对比页，好坏由人看了定。
聊天第二版：表情和动作节奏进入聊天。
头慢慢漂的视差两个角色都换上了，澄夏歪头时脖子跟着转。
九个画图模型的官方完整版在三台 GPU 机器上全部就位，每个文件逐一校验。
这些是方向，
还不代表已有合作。
有脸、有声音、会反应的 AI 陪伴角色，在网页里就能聊。
批量产出能动的 2D 角色，直接是游戏常用的 Spine 格式。
从一张设定图到能开口说话的形象，省掉建模和手工绑定。
品牌的原创角色，快速变成会说话、有表情的代言人。
完整权重、bf16 / fp16、官方步数、原分辨率。压缩、蒸馏、少步数的版本只能当标明的草稿。
她的行囊 · 我们的行囊
Qwen3/SenseVoice/Qwen3-TTS/CosyVoice2/流式语音会话
Spine 4.2 运行时/自研舞台/Next.js 16/React 19
视频老师/物理约束/视差分层
Qwen-Image-Edit-2511/FLUX.2 klein base/Qwen-Image-2.1/FLUX.1 Fill/Anima
See-through/Qwen-Image-Layered/SAM 3.1
RTX 4090/DGX × 2/持久任务队列/局域网同步、逐文件校验
这是我们正在冲的终点线：把设定图交给流水线，同一套代码、同一套参数，长出下一个会聊天的角色。
Aoi Studio turns a single illustration into a Spine 4.2 skeletal-animation character, then connects text and real-time voice. Swap in another character and the same code runs unchanged.
This clip is Summer running live in a web page. Not a single frame is hand-drawn animation.
Like a spatial photo
that breathes.
She is rendered frame by frame in the browser, not a pre-recorded animation.
As her head drifts, front and back layers move at different speeds, giving a flat illustration real depth.
Twintail's chat runs the very same parallax code.
One drawing, one sentence —
that is all her luggage.
A warm, lively, endlessly curious traveler who loves hearing about your day.
Fig. 2 The eyes in her skeleton, swapped frame by frame as her expression changes
One design sheet already holds four of her moods.
— from the design sheetTwintail and Summer come out of the same pipeline and the same code. Below, both stand idle, full body, recorded live from a web page.
The numbers differ because nothing is copied: the same code reads each character's own art and grows bones and meshes that belong to her.
Fig. 3 The eyes in Twintail's skeleton
An animated anime character traditionally takes several manual crafts: illustration, cutting parts, painting hidden areas, rigging, animating. We hand every one of them to code and models.
Body & capeForearms & handsFace & featuresFront hair & braid
Left: her, running live in the page. Above: eye parts from her skeleton.
Every number is read straight from the skeleton file. No estimates, no rounding up.
One step at a time,
every step on record.
Summer is born: midnight-blue hair, amber eyes, a travel cape and a compass hairpin.
The first commit: chat page, scripts and docs set off together.
Summer and Twintail move for the first time on pipeline-built skeletons, with living faces.
Every line's results land on one interactive comparison page. People, not metrics, decide what is good.
Chat v2: expressions and motion timing arrive in chat.
Both characters get the slow-drifting parallax; Summer's neck now turns with her head tilt.
Official full-precision builds of nine image models are in place on three GPU machines, every file checksum-verified.
These are directions,
not existing partnerships.
An AI companion with a face, a voice and reactions, right in the browser.
Animated 2D characters at scale, delivered in Spine, the format games already use.
From one design sheet to a persona that speaks, without modeling or manual rigging.
Turn a brand's original character into a talking, expressive ambassador, fast.
Full weights, bf16 / fp16, official step counts, native resolution. Compressed, distilled or few-step variants only ever appear as labeled drafts.
Her pack · our pack
Qwen3/SenseVoice/Qwen3-TTS/CosyVoice2/Streaming voice session
Spine 4.2 runtime/Custom stage/Next.js 16/React 19
Video teacher/Physics constraints/Parallax layers
Qwen-Image-Edit-2511/FLUX.2 klein base/Qwen-Image-2.1/FLUX.1 Fill/Anima
See-through/Qwen-Image-Layered/SAM 3.1
RTX 4090/DGX × 2/Persistent job queue/LAN sync, per-file checksums
This is the finish line we are racing toward: hand a design sheet to the pipeline and, with the same code and the same parameters, grow the next character who can talk.