01 UPPER BODY
CHARACTER CONSISTENCY
角 色 一 致 性C H A R A C T E R C O N S I S T E N C Y
V O L . 1 · 2 0 2 6
角 色 一 致 性C H A R A C T E R C O N S I S T E N C Y
V O L . 1 · 2 0 2 6
MORI
Theme & Variations
One face. Six plates.
Everything else changes.
Mori 不是一張圖，是一個可以重複召喚的人。換衣服、換光、換場景、換角度，臉不換。全書圖版皆為生成影像，非攝影。Mori is not one picture. She is a person who can be summoned again and again. Change the clothes, the light, the place, the angle — the face stays. Every plate in this volume is a generated image, not a photograph.
A study of identity through distance, posture and composition.
在景別、姿態與構圖之間，探索身份。Identity, explored between distance, posture and composition.
變化發生在人物之外。The change happens everywhere except her.
服裝、位置、鏡頭與節奏持續改變，而 Mori 保持不變。The clothes, the position, the camera and the tempo keep changing. Mori does not.
影片以六十秒作為尺度，讓同一個身份在不同視覺條件下反覆出現。Sixty seconds is the measure: one identity, made to reappear under one visual condition after another.
音樂的問題不是生成不出來，而是時間不夠準。The problem with music was never generating it, but generating it on time.
它可以有情緒，也可以有結構。It can carry a mood, and it can carry a structure.
但很難要求它剛好在第幾秒轉折、停下來，或替某一個畫面落點。What it cannot do is turn, stop, or land on a given second for a given shot.
所以不讓音樂決定時間。So the music was not allowed to decide the timing.
先把影片長度和主要切點固定，再用速度與段落去限制生成範圍，最後把音樂重新剪回畫面的節奏裡。Fix the length of the film and its main cuts first, narrow what can be generated with tempo and section structure, then cut the music back into the rhythm of the image.
當生成無法精準對齊，就先把它能移動的範圍縮小。When generation cannot align precisely, first narrow the range it is free to move in.
從既有的身體，From the body that already exists,
拼合出尚未被看見的部分。assemble the part that has not been seen.
沒有腿的位置，模型自然不會長出腿。解決它的不是再換一句 prompt，而是換一種方式：保留她已經存在的部分，再補上還不存在的部分。With no room for legs, the model was never going to grow any. What solved it was not another prompt but another method: keep the part of her that already exists, then add the part that does not.
比例不是靠 prompt 拉出來的。Proportion is not something a prompt can stretch.
當參考圖無法改變比例，就直接改變形變本身。When the reference cannot change the proportion, change the deformation itself.
修不好的地方，缺的不是更好的描述。Where a repair keeps failing, what is missing is not a better description.
先把結構定下來，肢體才有了可以遵循的形狀。Fix the structure first, and the limb finally has a shape to follow.
角度、姿勢、比例，都可以被重新指定。Angle, pose and proportion all become things that can be specified again.
兩個以上的她同時出現，prompt 很難說清楚誰該站在哪裡。Once more than one of her is in frame, a prompt cannot say clearly who should stand where.
先用 Blender 排好位置與運鏡，再交給影片模型。Block the positions and the camera move in Blender first, then hand it to the video model.
不需要真的走路，只要知道誰在什麼時候出現、往哪裡去。It does not need to walk properly. It only has to say who appears when, and which way they go.
第一版：臉留下來了，嘴也一起被記住了。First pass: the face carried over, and so did the mouth.
身份很穩定。Identity was stable.
服裝、取景、表情和嘴型也跟著固定下來。Wardrobe, framing, expression and mouth settled into place with it.
最明顯的是嘴——她總是微微張著。The mouth was the most obvious: always slightly parted.
直接要求閉嘴沒有用，畫面反而會被一起推動。Asking it to close did nothing; it moved the whole frame instead.
修正版：先讓資料裡真的存在另一種嘴型。Second pass: first make the other mouth actually exist in the data.
補進閉嘴的樣本，再把每張圖的嘴型寫進標註。Add closed-mouth samples, then write the mouth into every caption.
有了張嘴與閉嘴的對照，嘴才重新變成可以指定的東西。With open and closed to compare, the mouth became something you can ask for again.
沒有被分開標註的差異，最後都會被模型當成角色本身。Any difference the captions never separate ends up read as the character itself.
Turning visual ideas into generative workflows.
[email protected] @frameboundart全書圖版與短片皆為生成影像，非攝影。Every plate and the film are generated images, not photographs.MORI 為原創虛擬角色MORI is an original virtual character，不對應任何真實人物。, and corresponds to no real person.