角色連貫性
創作有明確辨識度角色嘅漫畫序列
用身份參考、當前服裝同逐格檢查，規劃角色一致性 AI 漫畫。了解當鏡頭、姿勢或場景改變時要檢查啲乜。
模型
呢個創作器會生成靜態圖片。漫畫同影片請繼續用連結工具；研究策略唔係可選生產模式。
圖片記錄
Google Vertex AI · gemini-3.1-flash-image · · 1280 × 859 px
原始生成提示詞
Use case: photorealistic-natural / product-mockup. Create a natural window-lit editorial PHOTOGRAPH of exactly TWO anatomically normal adult hands holding one landscape matte comic proof sheet above a real warm oak desk. Image1 supplied is the original adult botanist and mint robot reference sheet; image2 is the COMPLETE original SIX-PANEL greenhouse comic. The held paper must have landscape 3:2 proportions so Image2 fits WITHOUT rearrangement. Put the exact complete Image2 sequence on the held sheet as THREE columns by TWO rows, with a clean ivory outer margin: upper row arrival, pot on bench, gloved seed close-up; lower row watering, seedling reaction, final greenhouse view. Preserve all six distinct scenes in that order; DO NOT repeat watering, remove seed, reformat into two columns, add text or combine panels. Keep the original adult's teal jacket, coral scarf, dark curly hair, sand trousers and original mint round robot. Hold the paper by its two lower blank outside corners only; all artwork corners remain visible and hands cover no panels. The source character-reference sheet from Image1 rests flat farther back on the desk, partly visible beside the held print. Only the printed artwork is ink/watercolor illustration; hands, table, thin paper grain and room are a convincing photograph. No visible faces, booklet, screen, logos, titles, numbers, writing, fake UI, metrics or product packaging. Natural skin/five fingers, coherent optical perspective, subtle paper flex, quiet soft daylight. The completed hand-held story takes 75 percent of frame. Fictional lifestyle presentation only, never a customer photograph or research result. Landscape 3:2.
查看完整作品規劃具體結果
同一位植物學家，三種鏡頭距離
規劃一段溫室序列，包含建立場景、種植動作同特寫。成年植物學家喺鏡頭構圖轉換時保持短深色曲髮、青綠工作外套同珊瑚色頸巾；細小薄荷色機械人維持兩隻眼同一支天線。
要準備啲乜
- 一張清楚見到身份特徵嘅原創身份圖片。
- 溫室場景當前服裝同道具備註。
- 三個鏡頭簡述，標明動作同鏡頭距離。
圖片記錄
Google Vertex AI · gemini-3.1-flash-image · · 1280 × 859 px
原始生成提示詞
Use case: illustration-story. Original character reference sheet on warm ivory paper, editorial comic style with elegant clean dark ink lines, restrained watercolor washes and sophisticated readable shapes. Show the same fictional adult botanist, about 32, warm brown skin, short dark curly hair, teal work jacket over cream shirt, one coral scarf, sand trousers and simple brown ankle boots, in THREE clear full-body views: front, side and back, spaced evenly. Below or beside the adult show a small round mint-green helper robot in front and side view, with two dark circular eyes and exactly one short antenna. Leave comfortable margins, warm ivory background. Keep clothing color, hair, build and robot construction consistent between views. No writing, labels, logos, speech bubbles, charts, numbers, duplicated limbs, weapons, third-party IP or sexualized clothing. This is original illustration, not a technical accuracy or model benchmark claim. Landscape 3:2.
查看完整作品將身份同換衫分開處理
除下外套係故事狀態改變，唔係身份出錯。請記錄發生時間，咁之後嘅分鏡先可以同正確外觀比較。
揀一個有用嘅視角
後視圖需要輪廓同衣著剪裁；特寫就需要面部細節。應比較相關參考組，而唔係每個鏡頭都複製同一組選擇。
明確檢視互動
喺雙人分鏡入面，要逐個檢查每個面孔、服裝同相對位置。淨係角色數量正確，並唔足夠。
一致性其實係一組審核問題。角色應該保持可辨識，同時又可以自然地動作、轉身同變化。
一個實際可行嘅起點
定義身份同目前外觀
揀一個原創參考，並將穩定特徵同可改變嘅服裝、配件同故事狀態分開記錄。
將參考對應返鏡頭
面部特寫、全身動作同後視圖需要唔同嘅視覺資訊。應先用少量相關參考，而唔好以為圖片越多就一定越好。
一齊檢視序列
將分鏡並排比較。檢查出現咗邊個、著咗咩、企喺邊，同埋每個改動有冇被故事交代。
要檢查啲乜
圖片記錄
Google Vertex AI · gemini-3.1-flash-image · · 1280 × 859 px
原始生成提示詞
Use case: photorealistic-natural / product-mockup. Create a natural window-lit editorial PHOTOGRAPH of exactly TWO anatomically normal adult hands holding one landscape matte comic proof sheet above a real warm oak desk. Image1 supplied is the original adult botanist and mint robot reference sheet; image2 is the COMPLETE original SIX-PANEL greenhouse comic. The held paper must have landscape 3:2 proportions so Image2 fits WITHOUT rearrangement. Put the exact complete Image2 sequence on the held sheet as THREE columns by TWO rows, with a clean ivory outer margin: upper row arrival, pot on bench, gloved seed close-up; lower row watering, seedling reaction, final greenhouse view. Preserve all six distinct scenes in that order; DO NOT repeat watering, remove seed, reformat into two columns, add text or combine panels. Keep the original adult's teal jacket, coral scarf, dark curly hair, sand trousers and original mint round robot. Hold the paper by its two lower blank outside corners only; all artwork corners remain visible and hands cover no panels. The source character-reference sheet from Image1 rests flat farther back on the desk, partly visible beside the held print. Only the printed artwork is ink/watercolor illustration; hands, table, thin paper grain and room are a convincing photograph. No visible faces, booklet, screen, logos, titles, numbers, writing, fake UI, metrics or product packaging. Natural skin/five fingers, coherent optical perspective, subtle paper flex, quiet soft daylight. The completed hand-held story takes 75 percent of frame. Fictional lifestyle presentation only, never a customer photograph or research result. Landscape 3:2.
查看完整作品研究方向
工作流程背後嘅研究
用更少但更相關嘅參考
我哋正探索點樣比較唔同參考選擇，例如只用身份、加強人臉、同鏡頭匹配嘅組合。用較少參考得到更好結果，仍然只係待驗證嘅假設。
確保身份對應返正確人物
多角色研究會探討每個身份點樣保持同預期位置、服裝同比例相連，而唔會錯位去另一個角色。
適當評估人物同吉祥物
真實面孔同風格化送種子機械人需要唔同檢查。對非人角色，除咗影像相似度之外，仲要檢視整體輪廓、標誌性結構同人類判斷。
研究方向
角色一致性基準
當鏡頭、姿勢、衣著同場景改變時，角色有幾穩定仲可辨識？
我哋會點樣評估: 要分開評估身份、衣著、角色數量、故事一致性，同每張已接受分鏡嘅總成本。
研究方向
自適應參考選擇
邊種身份、全身、服裝或鏡頭視角參考，最有助某一個鏡頭？
我哋會點樣評估: 用相同故事同接受準則，去比較固定參考組同鏡頭對應組。
研究方向
多角色身份
喺互動過程中，每個角色都可唔可以保持正確面容、衣著、比例同位置？
我哋會點樣評估: 檢查每個鏡頭入面身份同位置嘅對應、角色數量同相對大小。
了解界線
- 參考圖同提示詞唔保證每一格都可以完全一致。
- 服裝、年齡、姿勢同風格改動，都需要清楚標明故事狀態，先可以公平評估。
- 研究問題同示範場景唔等於已發布嘅一致性分數或客戶成果。
初步研究只會為呢啲問題提供參考；並唔代表 CyaniModel 嘅結果。
常見問題
我嘅角色參考集應該包括啲乜？
先用已授權嘅身份圖，必要時加清晰全身視圖同當前服裝。如果計劃鏡頭需要第一張圖未有顯示嘅資訊，就再加側面或背面視圖。
係咪發晒所有參考圖會更好？
呢個係一個實驗性問題。額外圖片可能冇關，或者同當前服裝衝突。用同一份鏡頭簡報，比較一組較細但相關嘅圖片同一組較大圖片。
應該點樣評估動物或者吉祥物？
請檢視整體外形、比例、顏色同獨特結構。單靠人臉分數，未必可以證明一隻卡通動物或機械人係同一角色。
CyaniModel 會為每個角色都訓練一個模型嗎？
本指南唔代表提供角色專屬訓練服務。Character LoRA 同共用身份適配器都係研究選項，佢哋嘅準備成本同效益需要另外評估。
繼續你嘅故事
喺容許故事、姿勢同鏡頭變化嘅同時，保持角色可辨識。
繼續喺 LlamaGen 入面查看可用工具、模型選項、帳戶要求同積分，再進行生成。