形象資產

阿兔兔

多角度形象、LoRA 訓練素材、以及三個訓練階段的比對

⭐ 阿兔兔的造型一個像素都不能改 —— 那是樂樂真實抱著睡覺的那隻 Jellycat。 文字描述鎖不住質感，所以除了每次都餵參考圖，另外訓了一顆專屬的 LoRA。

唯一真相

阿兔兔實物照

實物照

基準

奶油白絨毛、長垂耳、草莓＋小白花耳襯、粉色小三角鼻、小黑豆眼、碎花腳掌、坐姿修長。

⛔ 所有生成都以這張為準。統一的方向永遠是「其他角色配合阿兔兔」，不是反過來。

LoRA 訓練素材（22 張）

每一張刻意換一個角度、一種光線、一個背景 —— 讓模型學到的是這隻玩偶本身， 不是某個特定的拍法。下面的說明就是餵給模型的標註。

atutubunny 的臉部特寫，正臉對著鏡頭，光從側面柔柔地照過來 close-up of atutubunny face, front view, soft side light
atutubunny 的臉部特寫，臉稍微往左轉，暖黃色的燈光，背景是花園 close-up of atutubunny face, three-quarter left, warm light, garden background
atutubunny 的臉部特寫，臉稍微往右轉，清冷的燈光，背景是藍色的 close-up of atutubunny face, three-quarter right, cool light, blue background
atutubunny 的頭部特寫，拍正側面，背後有光照過來，背景暗暗的 close-up of atutubunny head, side profile, backlit, dark background
atutubunny 的臉部特寫，抬頭往上看，頭頂有很亮的強光，背景是白的 close-up of atutubunny face, looking up, bright top light, white background
atutubunny 的臉部特寫，微微低著頭，晚上溫暖的檯燈光 close-up of atutubunny face, head tilted down, warm lamp light, night
atutubunny 的臉部特寫，臉部一半藏在陰影裡，側面的強光很有氣氛 close-up of atutubunny face, half in shadow, dramatic side light
atutubunny 的頭部特寫，從低角度往上拍，明亮的日光，背後是天空 close-up of atutubunny head, low angle, daylight, sky behind
atutubunny 的頭部特寫，從正上方往下拍的視角，背景是木地板 close-up of atutubunny head, high angle from above, wooden floor
atutubunny 耳朵布料細節的微距特寫，看得到纖維質感 macro close-up of atutubunny ear fabric detail
atutubunny 坐在攝影棚裡的全身正面照 full body of atutubunny sitting, front view, studio
atutubunny 坐在滿是雛菊的草地上，白天的自然光線 full body of atutubunny sitting in a daisy meadow, daylight
atutubunny 趴在地上的全身照，由正上方往下拍，透著清晨的陽光 full body of atutubunny lying flat, top-down view, morning light
atutubunny 靠在樹幹上的全身照，森林裡的陽光點點灑在身上 full body of atutubunny against a tree trunk, dappled forest light
atutubunny 坐著的半身照，正對著鏡頭，打著攝影棚的燈光 medium shot of atutubunny sitting, front view, studio light
atutubunny 坐在廚房桌上的半身照，臉轉向左邊 medium shot of atutubunny sitting on a table, turned left, kitchen
中景鏡頭，atutubunny 坐在草地上，大晴天 medium shot of atutubunny sitting on grass, sunny day
中景鏡頭，atutubunny 的背影 medium shot of atutubunny from behind, back view
中景鏡頭，atutubunny 側躺在毯子上，燈光昏黃溫暖 medium shot of atutubunny lying on its side on a blanket, dim warm light
中景鏡頭，atutubunny 靠在靠墊上，窗外透進來的光線 medium shot of atutubunny leaning on a cushion, window light
中景鏡頭，atutubunny 坐在秋天的落葉堆裡，傍晚夕陽光 medium shot of atutubunny sitting in autumn leaves, golden hour
中景鏡頭，atutubunny 坐在石階上，陰天 medium shot of atutubunny sitting on a stone step, overcast

訓練成果比對

三個階段各生同樣三張。老闆挑的是 epoch 6（1320 steps）—— 造型穩定又還沒過擬合。

epoch 2440 steps · 還抓不住耳襯的碎花
epoch 4880 steps · 造型對了，質感偏硬
epoch 6⭐ 選定 · 1320 steps
觸發詞atutubunny —— 每個 prompt 都要有，否則 LoRA 不作用
正式檔C:\ComfyUI\models\loras\atutu_SELECTED.safetensors（廣州）
訓練素材22 張 ＋ caption
目前用途畫面生成時鎖造型。影片生成仍以參考圖注入為主 —— ⛔ 圖對不代表影片會對，動態過程模型還是會自己改

其他角色

目前只有阿兔兔有專屬 LoRA。其餘角色靠定裝照做參考圖注入 —— 跨格的細節漂移（耳朵長度、嘴巴顏色）就是這樣來的，之後可能要比照辦理。