AI 漫劇〈雨夜狐廟〉
本機製作流程與提示詞
這支 48 秒的漫劇預告照製作順序分成六步,附上分鏡表、每一段生圖和生影片的提示詞、本機設定和實測時間
黃志弘 Leo
2026-10-25 發佈 · 3 人看過
〈雨夜狐廟〉是一支 48 秒的 AI 漫劇預告片。12 個鏡頭的畫面、對白和環境音,是我在自己的 Mac Studio 上用 MiniMax H3 生成的,配樂也在同一台電腦上用 ACE-Step 1.5 生。角色設定圖和每一鏡的第一格用 ChatGPT 網頁版生,系統提示音用 Gemini 的語音模型錄,這兩樣是雲端服務。跑模型、剪接和混音的程式,是 Claude Code 寫的。
每段提示詞上方都標了來源。標「原文」的,是當時真的送出去的內容;標「整理版」的,是事後依當時的做法整理成可以直接貼的一段,沒有再重跑一次。
雲端影片模型生 30 秒的牌價
2026-10-24 各家官方價目頁上,最新模型最高規格的牌價,換算成每秒和生一支 30 秒影片的價錢:
| 模型 | 規格 | 每秒 | 30 秒約 |
|---|---|---|---|
| Google Veo 3.1 | 4K,有聲音 | US$0.60 | NT$586 |
| ByteDance Seedance 2.5 | 1080p | US$0.5686 | NT$556 |
| Kling 3.0 | 4K,有聲音 | US$0.42 | NT$411 |
| MiniMax H3 雲端版 | 2K | US$0.13 | NT$127 |
價目來源:Google Cloud、BytePlus、Kling、MiniMax,2026-10-24 查。匯率 1 美元=32.58 台幣,臺灣銀行美金即期賣出,2026-10-24 09:25 牌價。都是牌價,沒算重生的次數,也沒算折扣
- Veo 3.1:價目頁寫的是「$0.60 / 1 count」,對照 Gemini API 價目頁,同一個數字標的是每秒,這裡照每秒算。一次最多生 8 秒
- Seedance 2.5:最高 1080p,價目是每百萬 tokens US$11.7,1080p 每秒 48,600 tokens,換算每秒 US$0.5686
- Kling 3.0:每秒 3 個 Unit,1 Unit 定價 US$0.14。更新的 Kling 4.0 只開放搶先體驗,價目頁上沒有價錢
- MiniMax H3 雲端版:官方頁沒寫輸出有沒有聲音。同一張價目表上 768P 是每秒 US$0.08,30 秒約 NT$78,這一檔的解析度跟我本機的輸出比較接近
一集漫劇要幾十個鏡頭,有的鏡頭還要重生好幾次。這支預告的影片和配樂改在自己的電腦上跑,生幾次都不用再按秒付錢;電費我沒有實測,沒算進來。
本機用的電腦、模型和生成時間
| 電腦 | Mac Studio,Apple M5 Max,64 GB 記憶體 |
|---|---|
| 影片和對白 | MiniMax H3 的首尾幀版本(fl2va),在 ComfyUI 裡跑。主模型用 unsloth 量化的 GGUF(Q8_0),文字編碼器用 Q4_K_M,加上官方的 8 步加速 LoRA |
| 配樂 | ACE-Step 1.5,XL turbo 模型出聲音,4B 語言模型作曲 |
| 角色和第一格 | ChatGPT 網頁版的生圖功能(雲端) |
| 系統提示音 | Gemini 語音模型 gemini-3.1-flash-tts-preview(雲端) |
| 剪接、字幕、混音 | ffmpeg、Whisper,程式由 Claude Code 寫 |
H3 生一鏡要多久:2026-10-22 第一批 10 鏡,73 幀(3.04 秒)的每鏡 11.6 到 13.5 分鐘,90 幀(3.75 秒)的每鏡 14.6 到 15.7 分鐘,10 鏡從凌晨 3:05 跑到 5:27,共 2 小時 22 分。2026-10-24 加做的結尾鏡頭用同樣的設定,73 幀 7.4 分鐘、90 幀 10.4 分鐘,比第一批快,原因我沒有查。記憶體用量最高約 33 GB,首尾幀那一鏡到 35 GB。
平均下來,成片 1 秒大約要 4 分鐘的本機運算。
第一步:寫分鏡,一鏡 3 到 4 秒
劇情是原創的:在公司加班的女生醒來,變成被流放的囚犯,和受傷的押送官差躲在雨夜的破廟裡。一隻巨大的九尾白狐找上門,她的掌心浮出金色符文,綁定了斬妖系統。結構參考一部 3D 古風漫劇的開場,順序是穿越、身分壓力、危機、金手指。
節奏也照那部的開場,一鏡 3 到 4 秒,每一鏡的鏡頭都慢慢推近。我量過它開頭 4 分鐘,每分鐘 15.2 刀,鏡頭長度的中位數是 2.95 秒。寫分鏡的時候就把長度定好:有台詞的鏡頭生 90 幀(3.75 秒),沒有台詞的生 73 幀(3.04 秒)。
| 鏡 | 長度 | 畫面 | 聲音 |
|---|---|---|---|
| 1 | 3.75 秒 | 雨夜山中的破廟遠景,閃電照亮屋頂,鏡頭推向廟門 | 內心獨白「我明明還在公司加班……」 |
| 2 | 3.75 秒 | 俯拍她躺在乾草上的臉,眼睛慢慢對焦,一滴雨落在臉頰 | 內心獨白「一睜眼,我成了流放的囚犯。」 |
| 3 | 3.04 秒 | 她把雙手舉到胸前,手腕上是生鏽的鐵鐐和短鐵鍊 | 鐵鍊聲、滴雨聲 |
| 4 | 3.75 秒 | 受傷的押送官差靠著紅柱坐在地上,抬頭看她 | 官差低聲說「別出聲。過來。」 |
| 5 | 3.04 秒 | 她半坐起身、轉頭看向廟門,閃電照亮半邊臉 | 雨聲、門外遠遠的低吼 |
| 6 | 3.04 秒 | 門外漆黑的雨幕裡,一雙金色豎瞳慢慢逼近 | 越來越大的低吼 |
| 7 | 3.75 秒 | 官差拄著長刀站起來,把刀整個拔出來 | 官差說「它聞到活人了。」、拔刀聲 |
| 8 | 3.04 秒 | 低角度仰拍,巨大的白狐從廟門探進頭來 | 風聲、白狐的喘息和低吼 |
| 9 | 3.04 秒 | 她的右手掌心浮出一圈金色古篆符文 | 上揚的鈴聲和能量聲,成品另外加上系統提示音 |
| 10 | 3.75 秒 | 她直視鏡頭,眼神從驚恐變成冷靜 | 內心獨白「好,那就試試看。」 |
| 11 | 3.04 秒 | 從背後跟拍,她衝出廟門跑進雨裡,白狐在前方 | 踩水的腳步、鐵鍊、雷聲、白狐低吼 |
| 12 | 3.75 秒 | 接著跑兩步、蹬地起跳,騰空撲向白狐(首尾幀) | 踩水起跳、風聲、雷聲 |
長度是 H3 生成的長度,剪接時第 9、10、12 鏡另外調過(見第六步)
寫分鏡的提示詞
整理版:依這支分鏡實際用到的規則整理,【題材】換成你的故事
幫我寫一支約 30 秒的漫劇開場分鏡,題材是【題材】,畫風是 3D 寫實 CG 古風玄幻。
1. 劇情原創,順序是穿越、身分壓力、危機、金手指
2. 一共 10 鏡,一鏡 3 到 4 秒,每一鏡的鏡頭都慢慢推近
3. 每一鏡寫:景別和角度、光線、人物動作、聲音(對白、內心獨白、環境音)
4. 台詞很少,一鏡最多一句;主角心裡的話寫成內心獨白
5. 有台詞的鏡頭標 3.75 秒,沒有台詞的標 3.04 秒
最後列出要先做的設定圖:主角的臉、主角的身體和服裝三視圖、每一個配角、主要場景。第二步:在同一個對話裡生出所有圖
角色設定圖 4 張、每一鏡的第一格 11 張,全部在同一個 ChatGPT 對話裡依序生。後面每一張都看得到前面的臉和服裝,換到下一鏡,角色才不會換一張臉。第一張花了約 6 分鐘,之後每張約 40 到 60 秒。
提示詞有三個固定寫法。每一張開頭都寫明是設定圖或劇照、橫式 16:9;結尾都貼同一段畫風描述,一字不改;每一鏡的第一格用「上面那位女主角(臉照第一張臉部設定、服裝照第二張)」「照第三張設定」這種寫法,指回前面的設定圖。設定圖都用淺灰背景、柔和棚光,破廟的場景圖裡不放人。


以下 15 段都是原文,第 1 到 14 張 2026-10-22、第 15 張 2026-10-24 送進同一個 ChatGPT 對話
1. 女主角的臉部設定(第一張,定臉)
角色設定圖(橫式 16:9):同一位約二十歲的古代年輕女子的臉部設定,一張圖內並排正面、3/4 側面、側面三個頭像,大而清楚。鵝蛋臉、眉形細長、杏眼眼尾微挑、鼻樑秀挺、薄唇,黑色長髮半披、用一根舊木簪隨意挽起,幾縷濕髮貼在臉頰,臉上有一點泥痕。表情冷靜、帶點警覺。乾淨淺灰背景、柔和棚光,只畫頭部不畫身體。畫風:3D 寫實 CG 古風玄幻,像高品質國產 3D 動畫劇集或遊戲過場動畫的畫面,皮膚有細緻毛孔質感、髮絲分明、布料有重量,電影級光影。不要任何文字、字幕、浮水印。2. 女主角的身體和服裝三視圖
角色設定圖(橫式 16:9):上一張那位女子的身體與服裝三視圖(正面、側面、背面),A-pose 站姿,頭部不要畫出來(畫面在脖子處截斷)。身材纖瘦,穿粗麻灰白色囚衣,衣角破損、沾泥水,腰間繫粗麻繩,雙手手腕戴一副生鏽的鐵鐐、中間連一段短鐵鍊,赤腳穿破舊草鞋。乾淨淺灰背景、柔和棚光。畫風:3D 寫實 CG 古風玄幻,像高品質國產 3D 動畫劇集或遊戲過場動畫的畫面,皮膚有細緻毛孔質感、髮絲分明、布料有重量,電影級光影。不要任何文字、字幕、浮水印。3. 押送官差
角色設定圖(橫式 16:9):另一個角色,約三十歲的古代押送官差。左半邊是臉部正面與 3/4 側面兩個大頭像:劍眉、眼神銳利疲憊、短鬍渣、髮髻用黑布束起,嘴角有一道乾掉的血痕,臉色蒼白。右半邊是全身正面 A-pose:深黑紅色的衙役勁裝外罩舊皮甲,腰佩一把長刀,左肩衣服撕裂纏著布條。乾淨淺灰背景、柔和棚光。畫風:3D 寫實 CG 古風玄幻,像高品質國產 3D 動畫劇集或遊戲過場動畫的畫面,皮膚有細緻毛孔質感、髮絲分明、布料有重量,電影級光影。不要任何文字、字幕、浮水印。4. 破廟場景
場景設定圖(橫式 16:9):雨夜山中的破舊土地廟內部,3/4 斜角度,看得到廟門、左右兩側與後方。屋頂破了一個大洞,雨水從洞口落下在地上積成水窪;正中央一尊斑駁傾倒的土地公石像;右側一根粗大的紅漆木柱(當參照物);地上散落乾草、翻倒的燭台,唯一光源是角落一盞快熄的油燈(暖橘光)加上門外偶爾的冷藍閃電。廟裡沒有任何人。畫風:3D 寫實 CG 古風玄幻,像高品質國產 3D 動畫劇集或遊戲過場動畫的畫面,皮膚有細緻毛孔質感、髮絲分明、布料有重量,電影級光影。不要任何文字、字幕、浮水印。5. 第 1 鏡第一格:破廟遠景
劇照(橫式 16:9):雨夜山中破舊土地廟的外觀遠景,大雨斜打,天空一道閃電照亮廟的輪廓,廟門半開透出一點暖橘色油燈光,周圍是黑色的松林與泥濘山路。沒有人物。畫風:3D 寫實 CG 古風玄幻,像高品質國產 3D 動畫劇集或遊戲過場動畫的畫面,皮膚有細緻毛孔質感、髮絲分明、布料有重量,電影級光影。不要任何文字、字幕、浮水印。6. 第 2 鏡第一格:醒來
劇照(橫式 16:9):上面那位女主角(臉照第一張臉部設定、服裝照第二張)仰躺在破廟的乾草上,臉部大特寫、俯拍,雙眼剛剛睜開、瞳孔還有點失焦,一滴雨水從屋頂落在她臉頰上,濕髮貼在額頭。冷藍的閃電光從上方打下,臉的一側有暖橘油燈光。畫風:3D 寫實 CG 古風玄幻,像高品質國產 3D 動畫劇集或遊戲過場動畫的畫面,皮膚有細緻毛孔質感、髮絲分明、布料有重量,電影級光影。不要任何文字、字幕、浮水印。7. 第 3 鏡第一格:鐵鐐特寫
劇照(橫式 16:9):女主角雙手的特寫,她剛把手舉到胸前,手腕上生鏽的鐵鐐和短鐵鍊清楚可見,手指沾泥、指甲縫有泥,背景是模糊的乾草與破廟地面,暖橘油燈光從側面照來,淺景深。畫風:3D 寫實 CG 古風玄幻,像高品質國產 3D 動畫劇集或遊戲過場動畫的畫面,皮膚有細緻毛孔質感、髮絲分明、布料有重量,電影級光影。不要任何文字、字幕、浮水印。8. 第 4 鏡第一格:官差
劇照(橫式 16:9):中景,那位受傷的押送官差(照第三張設定)背靠破廟右側的紅漆木柱坐在地上,一手按著左肩傷口、一手搭在膝上的長刀刀鞘,抬頭看向畫面左側(女主角方向),神情嚴厲。屋頂破洞的雨水在他前方落下,油燈暖光從左側照在他臉上。畫風:3D 寫實 CG 古風玄幻,像高品質國產 3D 動畫劇集或遊戲過場動畫的畫面,皮膚有細緻毛孔質感、髮絲分明、布料有重量,電影級光影。不要任何文字、字幕、浮水印。9. 第 5 鏡第一格:轉頭看廟門
劇照(橫式 16:9):中近景,女主角半坐起身,雙手戴著鐵鐐撐在乾草上,轉頭望向畫面右側的廟門方向,表情警覺、屏住呼吸。背景虛化的廟內,冷藍閃電光從門口照在她半邊臉上。畫風:3D 寫實 CG 古風玄幻,像高品質國產 3D 動畫劇集或遊戲過場動畫的畫面,皮膚有細緻毛孔質感、髮絲分明、布料有重量,電影級光影。不要任何文字、字幕、浮水印。10. 第 6 鏡第一格:門外的金色豎瞳
劇照(橫式 16:9):從廟內往半開的廟門外看,門外是漆黑的雨幕,黑暗中只看得見一雙巨大的金色豎瞳在發光,離地很高,雨絲被金光照亮。門框兩側的破紙燈籠在風中晃。沒有人物。畫風:3D 寫實 CG 古風玄幻,像高品質國產 3D 動畫劇集或遊戲過場動畫的畫面,皮膚有細緻毛孔質感、髮絲分明、布料有重量,電影級光影。不要任何文字、字幕、浮水印。11. 第 7 鏡第一格:官差拔刀
劇照(橫式 16:9):中近景,受傷的官差用長刀撐地正要站起來,刀已出鞘一半,刀身反射門外的冷光,他咬著牙、額頭冒汗,盯著畫面右側的廟門。油燈暖光從下方照亮他的下巴。畫風:3D 寫實 CG 古風玄幻,像高品質國產 3D 動畫劇集或遊戲過場動畫的畫面,皮膚有細緻毛孔質感、髮絲分明、布料有重量,電影級光影。不要任何文字、字幕、浮水印。12. 第 8 鏡第一格:白狐探頭進廟
劇照(橫式 16:9):低角度仰拍廟門口,一隻比門還高的巨大白狐從雨幕中探進半個身子,九條尾巴的影子映在門外,白毛被雨打濕、金色豎瞳俯視廟內,強風把油燈的火焰吹得斜倒,乾草被吹起。畫風:3D 寫實 CG 古風玄幻,像高品質國產 3D 動畫劇集或遊戲過場動畫的畫面,皮膚有細緻毛孔質感、髮絲分明、布料有重量,電影級光影。不要任何文字、字幕、浮水印。13. 第 9 鏡第一格:掌心符文
劇照(橫式 16:9):特寫,女主角攤開的右手掌心,掌心浮現一圈發光的金色古篆符文陣,金光往上照亮她的手腕鐵鐐和下巴,周圍飄起細小的金色光點,背景是虛化的昏暗廟內。畫風:3D 寫實 CG 古風玄幻,像高品質國產 3D 動畫劇集或遊戲過場動畫的畫面,皮膚有細緻毛孔質感、髮絲分明、布料有重量,電影級光影。不要任何文字、字幕、浮水印。14. 第 10 鏡第一格:直視鏡頭
劇照(橫式 16:9):女主角臉部近景,正面直視鏡頭,掌心的金光從下方照亮她的臉,眼神從驚恐變成冷靜決斷,嘴角微微收緊,濕髮被風吹動,背景是虛化的巨大白色狐影。畫風:3D 寫實 CG 古風玄幻,像高品質國產 3D 動畫劇集或遊戲過場動畫的畫面,皮膚有細緻毛孔質感、髮絲分明、布料有重量,電影級光影。不要任何文字、字幕、浮水印。15. 第 11 鏡第一格:衝出廟門(2026-10-24 接著生)
劇照(橫式 16:9):從破廟內往廟門外看的背後視角,上面那位女主角(臉照第一張臉部設定、服裝照第二張:灰白破囚衣、腰間麻繩、雙腕生鏽鐵鐐連一段短鐵鍊、草鞋)正從半開的廟門衝出去、一腳已經踏進門外的大雨裡,畫面只看得到她的背影和半張側臉,濕髮和破衣角往後飛揚,兩腕之間的短鐵鍊甩起來,右手掌心還亮著金色符文光。門外漆黑的雨夜庭院裡,就是前面探進廟門的那隻巨大白狐(濕透的灰白長毛、金色豎瞳、身後好幾條尾巴),比她高好幾倍,在雨幕中俯身盯著她,閃電把斜打的雨絲照亮。廟內暖橘油燈光從她背後打出輪廓光,門外是冷藍色。畫風:3D 寫實 CG 古風玄幻,像高品質國產 3D 動畫劇集或遊戲過場動畫的畫面,皮膚有細緻毛孔質感、髮絲分明、布料有重量,電影級光影。不要任何文字、字幕、浮水印。第三步:每一格丟進本機的 MiniMax H3
每一鏡的第一格丟進本機的 MiniMax H3,畫面、對白和環境音一次生出來。這支用的設定:
- 只給第一格(首幀模式),最後一鏡另外給最後一格,見第四步
- 1312x736,24 fps
- 有台詞 90 幀(3.75 秒),沒有台詞 73 幀(3.04 秒);H3 的幀數要是 17 的倍數加 5
- 官方 8 步加速 LoRA,跑 8 步,取樣器 res_multistep、排程 simple,cfg 1,不寫負面提示詞
- seed 固定 7
提示詞照官方提示詞指南的格式分成四段:第一句說第一格(<Picture 1>)對在第 0 秒;integrated_multimodal_description 寫畫面、動作和台詞,台詞用 <d>[Chinese] …</d> 包起來;overall_soundscape 寫環境音;non_diegetic_music 是配樂,這裡都寫 N/A,配樂最後另外做。
每一鏡開頭的畫風句和場景句一字不改。女主角出現的鏡頭,再貼同一段她的身分描述:年紀、臉型、眉眼、髮型和木簪、囚衣、麻繩、鐵鐐;官差也有他自己的一段。生圖的時候,角色靠同一個對話維持;生影片的時候,每一鏡是分開跑的,模型只看得到這一鏡的提示詞和第一格,所以每一鏡都要重新把她講一遍。
原文:每一鏡都一樣的開頭兩句
Realistic 3D CG Chinese xianxia animated series style, like a high-end game cinematic, detailed skin texture and individual hair strands, heavy wet fabric. Rainy night inside and around a ruined mountain shrine, lit by a single warm orange oil lamp and cold blue lightning flashes from outside.原文:女主角的身分描述,第 2、5、10、11、12 鏡逐字相同
a young woman of about twenty with an oval face, long thin brows, almond eyes with slightly lifted outer corners, black hair half-down and loosely pinned with an old wooden hairpin, a few wet strands stuck to her cheeks, a smear of mud on her face, wearing a torn grey-white coarse linen prisoner robe tied with a hemp rope, rusty iron shackles joined by a short chain on both wrists女主角心裡的話,寫成「Her lips stay closed. Her inner thoughts are heard as a calm young female voice-over」,環境音那一段再寫一次「her lips do not move」。三句內心獨白用 Whisper 轉回文字,字全對;第 10 鏡她的嘴巴一直閉著,第 2 鏡中段嘴巴還是張開了一下,像在喘氣。官差的兩句對白要對嘴,就寫「His lips move clearly with each word」,兩鏡都對上了嘴型,字也全對(「它」被轉成同音的「他」)。
以下 11 段都是原文,第 1 到 10 鏡 2026-10-22、第 11 鏡 2026-10-24 送進本機 H3
第 1 鏡:破廟遠景,內心獨白
For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.
integrated_multimodal_description: [Shot 1] Realistic 3D CG Chinese xianxia animated series style, like a high-end game cinematic, detailed skin texture and individual hair strands, heavy wet fabric. Rainy night inside and around a ruined mountain shrine, lit by a single warm orange oil lamp and cold blue lightning flashes from outside. The shot begins exactly from <Picture 1>: a ruined mountain shrine seen from far away in heavy slanted rain, half-open doors leaking warm orange lamplight, black pine forest and a muddy path around it. The camera slowly pushes in toward the shrine door. A bright lightning flash lights up the sky and the shrine roof for an instant, then darkness returns. A calm young female voice-over, the inner thoughts of the heroine, low and slightly breathless, in Mandarin with a soft Taiwanese accent (S1), says: <d>[Chinese] 我明明還在公司加班……</d> No one is visible in the frame.
overall_soundscape: Heavy rain on leaves and roof tiles, a close crack of thunder after the flash, wind through the pines. Only the one female voice-over speaks.
non_diegetic_music: N/A第 2 鏡:醒來,內心獨白(嘴巴閉著)
For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.
integrated_multimodal_description: [Shot 1] Realistic 3D CG Chinese xianxia animated series style, like a high-end game cinematic, detailed skin texture and individual hair strands, heavy wet fabric. Rainy night inside and around a ruined mountain shrine, lit by a single warm orange oil lamp and cold blue lightning flashes from outside. The shot begins exactly from <Picture 1>: an overhead extreme close-up of a young woman of about twenty with an oval face, long thin brows, almond eyes with slightly lifted outer corners, black hair half-down and loosely pinned with an old wooden hairpin, a few wet strands stuck to her cheeks, a smear of mud on her face, wearing a torn grey-white coarse linen prisoner robe tied with a hemp rope, rusty iron shackles joined by a short chain on both wrists, lying on her back on dry straw. Her eyes are already open and unfocused; she blinks twice, her pupils slowly focus, and a drop of rainwater falls from the broken roof onto her cheek and runs toward her ear. Her brows draw together. The camera pushes in very slowly. Her lips stay closed. Her inner thoughts are heard as a calm young female voice-over, low and slightly breathless, in Mandarin with a soft Taiwanese accent (S1): <d>[Chinese] 一睜眼,我成了流放的囚犯。</d>
overall_soundscape: Rain dripping through the roof onto straw, distant thunder, her quiet breathing. Only the one female voice-over speaks; her lips do not move.
non_diegetic_music: N/A第 3 鏡:鐵鐐,沒有台詞
For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.
integrated_multimodal_description: [Shot 1] Realistic 3D CG Chinese xianxia animated series style, like a high-end game cinematic, detailed skin texture and individual hair strands, heavy wet fabric. Rainy night inside and around a ruined mountain shrine, lit by a single warm orange oil lamp and cold blue lightning flashes from outside. The shot begins exactly from <Picture 1>: a close-up of the woman's hands raised in front of her chest, rusty iron shackles on both wrists joined by a short chain, muddy fingers. She slowly turns her wrists, and the chain swings and clinks; her fingers tremble slightly. The camera pushes in slowly toward the shackles with a shallow depth of field.
overall_soundscape: The metal chain clinks against the shackles, rain dripping onto straw, a faint flicker of the oil lamp. No speech.
non_diegetic_music: N/A第 4 鏡:官差說話(對嘴)
For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.
integrated_multimodal_description: [Shot 1] Realistic 3D CG Chinese xianxia animated series style, like a high-end game cinematic, detailed skin texture and individual hair strands, heavy wet fabric. Rainy night inside and around a ruined mountain shrine, lit by a single warm orange oil lamp and cold blue lightning flashes from outside. The shot begins exactly from <Picture 1>: a medium shot of a man of about thirty, an ancient escort officer with sharp tired eyes under strong brows, short stubble, hair tied in a topknot with black cloth, a dried trace of blood at the corner of his mouth, a pale face, wearing a dark black-and-red officer uniform under worn leather armour, a long sabre, his torn left shoulder bound with cloth, sitting on the floor with his back against a thick red-lacquered pillar, one hand pressed to his wounded left shoulder, the other resting on the sheathed sabre across his knees. Rain falls through the broken roof in front of him. He lifts his eyes toward the left of the frame, where the woman is, and speaks low and urgent, a hoarse deep male voice in Mandarin (S1): <d>[Chinese] 別出聲。過來。</d> His lips move clearly with each word, then close. He tilts his head slightly toward himself as a signal. The camera pushes in slowly.
overall_soundscape: Rain through the broken roof, distant thunder. Only the man speaks, in a low hoarse whisper.
non_diegetic_music: N/A第 5 鏡:轉頭看廟門
For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.
integrated_multimodal_description: [Shot 1] Realistic 3D CG Chinese xianxia animated series style, like a high-end game cinematic, detailed skin texture and individual hair strands, heavy wet fabric. Rainy night inside and around a ruined mountain shrine, lit by a single warm orange oil lamp and cold blue lightning flashes from outside. The shot begins exactly from <Picture 1>: a medium close-up of a young woman of about twenty with an oval face, long thin brows, almond eyes with slightly lifted outer corners, black hair half-down and loosely pinned with an old wooden hairpin, a few wet strands stuck to her cheeks, a smear of mud on her face, wearing a torn grey-white coarse linen prisoner robe tied with a hemp rope, rusty iron shackles joined by a short chain on both wrists, half sitting up on the straw, her shackled hands pressing into it. She holds her breath and slowly turns her head toward the shrine door on the right of the frame; a flash of cold blue lightning from the doorway lights half of her face, and her eyes widen slightly. Her lips stay closed. The camera pushes in slowly.
overall_soundscape: Rain, wind pushing the door, a low distant animal growl from outside. No speech.
non_diegetic_music: N/A第 6 鏡:門外的金色豎瞳
For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.
integrated_multimodal_description: [Shot 1] Realistic 3D CG Chinese xianxia animated series style, like a high-end game cinematic, detailed skin texture and individual hair strands, heavy wet fabric. Rainy night inside and around a ruined mountain shrine, lit by a single warm orange oil lamp and cold blue lightning flashes from outside. The shot begins exactly from <Picture 1>: looking from inside the shrine out through the half-open door into pitch-black rain. In the darkness a pair of huge glowing golden slit-pupil eyes, high above the ground, slowly blink once and draw closer. Rain streaks are lit gold by the eyes. Torn paper lanterns on both sides of the door swing in the wind. The camera pushes in slowly toward the eyes. No people are visible.
overall_soundscape: A deep rumbling growl that rises in volume, rain, the lanterns creaking in the wind. No speech.
non_diegetic_music: N/A第 7 鏡:官差拔刀(對嘴)
For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.
integrated_multimodal_description: [Shot 1] Realistic 3D CG Chinese xianxia animated series style, like a high-end game cinematic, detailed skin texture and individual hair strands, heavy wet fabric. Rainy night inside and around a ruined mountain shrine, lit by a single warm orange oil lamp and cold blue lightning flashes from outside. The shot begins exactly from <Picture 1>: a medium close-up of a man of about thirty, an ancient escort officer with sharp tired eyes under strong brows, short stubble, hair tied in a topknot with black cloth, a dried trace of blood at the corner of his mouth, a pale face, wearing a dark black-and-red officer uniform under worn leather armour, a long sabre, his torn left shoulder bound with cloth, pushing himself up from the floor with his sabre, the blade already half drawn and reflecting cold light from the door. He grits his teeth, sweat on his forehead, eyes fixed on the door to the right of the frame, and says through his teeth, a hoarse deep male voice in Mandarin (S1): <d>[Chinese] 它聞到活人了。</d> His lips move with each word. Then he draws the blade fully out with a sharp metallic ring. The camera pushes in slowly.
overall_soundscape: The ring of a sabre being drawn, rain, a deep growl from outside. Only the man speaks.
non_diegetic_music: N/A第 8 鏡:白狐探頭進廟
For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.
integrated_multimodal_description: [Shot 1] Realistic 3D CG Chinese xianxia animated series style, like a high-end game cinematic, detailed skin texture and individual hair strands, heavy wet fabric. Rainy night inside and around a ruined mountain shrine, lit by a single warm orange oil lamp and cold blue lightning flashes from outside. The shot begins exactly from <Picture 1>: a low-angle shot of the shrine doorway. A giant white fox, taller than the door, pushes its head and shoulders in from the rain, its wet white fur dripping, golden slit-pupil eyes looking down into the shrine. Wind from outside blows the oil-lamp flame sideways and lifts straw into the air. The fox slowly lowers its head toward the camera. The camera slowly tilts up.
overall_soundscape: Howling wind, the fox's heavy breathing and low growl, straw rustling, rain. No speech.
non_diegetic_music: N/A第 9 鏡:掌心符文
For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.
integrated_multimodal_description: [Shot 1] Realistic 3D CG Chinese xianxia animated series style, like a high-end game cinematic, detailed skin texture and individual hair strands, heavy wet fabric. Rainy night inside and around a ruined mountain shrine, lit by a single warm orange oil lamp and cold blue lightning flashes from outside. The shot begins exactly from <Picture 1>: a close-up of the woman's open right palm. A ring of glowing golden ancient seal-script characters spins slowly on her palm and grows brighter, lighting her shackled wrist and her chin from below; tiny golden sparks rise and drift upward. Her fingers tense. The camera pushes in slowly toward the palm.
overall_soundscape: A rising shimmering chime, a soft humming of energy, rain fading into the background. No speech.
non_diegetic_music: N/A第 10 鏡:直視鏡頭,內心獨白(嘴巴閉著)
For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.
integrated_multimodal_description: [Shot 1] Realistic 3D CG Chinese xianxia animated series style, like a high-end game cinematic, detailed skin texture and individual hair strands, heavy wet fabric. Rainy night inside and around a ruined mountain shrine, lit by a single warm orange oil lamp and cold blue lightning flashes from outside. The shot begins exactly from <Picture 1>: a close-up of a young woman of about twenty with an oval face, long thin brows, almond eyes with slightly lifted outer corners, black hair half-down and loosely pinned with an old wooden hairpin, a few wet strands stuck to her cheeks, a smear of mud on her face, wearing a torn grey-white coarse linen prisoner robe tied with a hemp rope, rusty iron shackles joined by a short chain on both wrists, facing the camera, golden light from her palm lighting her face from below, wet hair moving in the wind, the blurred white shape of the giant fox behind her. Her frightened eyes steady and her jaw tightens; she lifts her eyes to look straight into the camera. Her lips stay closed. Her inner thoughts are heard as a calm young female voice-over, now steady and cold, in Mandarin with a soft Taiwanese accent (S1): <d>[Chinese] 好,那就試試看。</d> The camera pushes in slowly.
overall_soundscape: Wind, rain, the fading golden hum, a low growl behind her. Only the one female voice-over speaks; her lips do not move.
non_diegetic_music: N/A第 11 鏡:衝出廟門
For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.
integrated_multimodal_description: [Shot 1] Realistic 3D CG Chinese xianxia animated series style, like a high-end game cinematic, detailed skin texture and individual hair strands, heavy wet fabric. Rainy night inside and around a ruined mountain shrine, lit by a single warm orange oil lamp and cold blue lightning flashes from outside. The shot begins exactly from <Picture 1>: seen from behind inside the shrine doorway, a young woman of about twenty with an oval face, long thin brows, almond eyes with slightly lifted outer corners, black hair half-down and loosely pinned with an old wooden hairpin, a few wet strands stuck to her cheeks, a smear of mud on her face, wearing a torn grey-white coarse linen prisoner robe tied with a hemp rope, rusty iron shackles joined by a short chain on both wrists, is charging out through the shrine door into the heavy rain, a ring of glowing golden seal-script characters on her right palm. Outside in the dark rainy courtyard, the giant white fox with wet grey-white fur, golden slit-pupil eyes and many tails looms far taller than her and stares down at her. She sprints forward out of the door and splashes through the puddles, her wet hair and torn robe flying back, the short chain between her wrists swinging. The camera follows close behind her through the doorway, moving forward fast. A lightning flash lights up the fox and the falling rain.
overall_soundscape: Heavy rain, her fast footsteps splashing through puddles, the chain rattling, a rolling crack of thunder, the fox's deep growl. No speech.
non_diegetic_music: N/A第四步:結尾接成一鏡到底
預告片原本的結尾,是她在廟外正面對著白狐擺出架勢的一鏡。後來改成一鏡到底:第 11 鏡衝出廟門之後不切鏡頭,她接著跑、起跳、騰空撲向白狐,碰到之前畫面漸暗,接著出劇名。
做法是把第 11 鏡的最後一格,當成新一鏡的第一格:
- 從第 11 鏡的影片取出最後一格(第 72 格,1312x736)
- 回到同一個 ChatGPT 對話,把這一格當附件傳上去,請它畫一秒多之後、她已經騰空的背影,機位、光線和背景都照附件
- H3 換成首尾幀模式:第一格是第 11 鏡的最後一格,最後一格是 ChatGPT 生的騰空背影,90 幀,其他設定不變,跑了約 10 分鐘
- 接起來:H3 從靜止的畫面起步,第 12 鏡的前 3 格幾乎不動,直接接會像頓了一下。所以第 11 鏡的 73 格全部用上,第 12 鏡從第 3 格接。接縫那一步,相鄰兩格的平均像素差是 11.4,第 11 鏡自己跑步時每一步是 9.7 到 12.2,看起來就是她跑的下一步
- 慢動作:從第 12 鏡第 49 格起,用 ffmpeg 的 minterpolate 補幀,8 格之內從正常速度慢到 1/2.5,每一格都是補出來的新畫面。最後 1 秒漸暗到全黑,停 6 格黑畫面,再出劇名

16. 第 12 鏡的最後一格:騰空的背影
原文:2026-10-24 附上第 11 鏡最後一格送進同一個 ChatGPT 對話。第一次附圖沒有傳成功,生出一張側面的構圖,所以第二次開頭多了一句
請完全以這則訊息的附件為準(上一張側面角度的不算)。附件是上一個鏡頭(女主角從破廟衝出去、在雨夜庭院裡朝白狐跑)的最後一格。請畫同一個鏡頭一秒多之後的畫面(劇照,橫式 16:9):機位、角度、鏡頭焦距、光線和背景都完全照附件,一樣從她背後低角度跟拍,左邊是亮著暖燈的廟宇建築和石階,右邊是石牆,前方雨幕中是同一隻巨大白狐(濕透的灰白長毛、金色豎瞳、身後九條尾巴)。唯一的差別:同一位女主角(黑色長髮半披、用舊木簪挽起,灰白破囚衣,腰間麻繩,雙腕生鏽鐵鐐連一段短鐵鍊)已經從積水的泥地上奮力躍起,整個人騰空在半空中撲向白狐,仍然是背影,雙腳離地、膝蓋彎起,濕髮和破衣下擺往後飛揚,兩腕之間的鐵鍊甩起來,右手掌心的金色符文光朝向前方。她離白狐還有一大段距離,還沒有碰到。白狐壓低頭、張口齜牙迎向她。大雨斜打,一道閃電照亮天空。畫風:3D 寫實 CG 古風玄幻,像高品質國產 3D 動畫劇集或遊戲過場動畫的畫面,皮膚有細緻毛孔質感、髮絲分明、布料有重量,電影級光影。不要任何文字、字幕、浮水印。第 12 鏡:起跳騰空(首尾幀)
原文:2026-10-24 送進本機 H3,同時給第一格和最後一格
For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced, and at the end of the target video <Picture 2> (the last frame) is fully referenced.
integrated_multimodal_description: [Shot 1] Realistic 3D CG Chinese xianxia animated series style, like a high-end game cinematic, detailed skin texture and individual hair strands, heavy wet fabric. Rainy night inside and around a ruined mountain shrine, lit by a single warm orange oil lamp and cold blue lightning flashes from outside. The shot begins exactly from <Picture 1>: seen from low behind in the rainy courtyard outside the shrine, a young woman of about twenty with an oval face, long thin brows, almond eyes with slightly lifted outer corners, black hair half-down and loosely pinned with an old wooden hairpin, a few wet strands stuck to her cheeks, a smear of mud on her face, wearing a torn grey-white coarse linen prisoner robe tied with a hemp rope, rusty iron shackles joined by a short chain on both wrists, is sprinting away from the camera toward the giant white fox with wet grey-white fur, golden slit-pupil eyes and nine tails that looms ahead in the rain, a golden glow on her right palm. She takes two more running strides, splashing through the puddles, then plants her foot and leaps high into the air toward the fox, knees bent, wet hair and torn robe streaming back, the short chain between her wrists swinging, her glowing right palm thrust forward. The fox lowers its head and bares its fangs at her. The camera keeps following from low behind her and tilts up slightly as she rises. A lightning flash lights the sky. She is still far from the fox and does not touch it. The shot ends exactly on <Picture 2>: she is airborne in mid-leap, seen from behind.
overall_soundscape: Heavy rain, her splashing footsteps, a sharp splash as she pushes off the ground, a rush of wind as she rises, the chain rattling, a crack of thunder, the fox's growl. No speech.
non_diegetic_music: N/A補幀的參數
這支用的 ffmpeg minterpolate 參數(程式裡寫成 24 x 8 fps):先把影片補成 8 倍密度,再用程式照速度曲線取格
minterpolate=fps=192:mi_mode=mci:mc_mode=aobmc:me_mode=bidir:vsbmc=1:scd=none第五步:配樂在本機生,最重的一下對準起跳
配樂先研究再生成。我先量參考那部漫劇開頭 4 分鐘的聲音:用 demucs 把人聲和配樂分開,用 librosa 每 0.5 秒量一次音量和頻段,畫成一條強弱曲線,再用 PANNs 標出每一秒是什麼聲音。量出來有這幾個做法:
- 開場第一下就是一記重擊
- 第一句旁白開口前,配樂被拉低約 20 dB
- 對白的時候,人聲比配樂大,中位數約 10.8 dB
- 緊張的段落主要是低頻的脈動和嗡鳴,旋律很少
- 最重的一下之前,先抽到幾乎沒有聲音
- 前 4 分鐘幾乎量不到民族樂器
接著照這支的分鏡,逐鏡寫出情緒、配樂在做什麼、緊張點落在哪幾秒,做成一條目標曲線。配樂用 ACE-Step 1.5 在本機生,全部設 E 小調、128 BPM、純配樂,一批 2 首 36 秒的候選約 70 秒生完。一共生了 12 個候選,有整首一次生的、分成兩段生的、只重畫尾段的,每一個都量音量和起音強度,跟分鏡的時間軸對照,挑最貼目標曲線的組合:前段用一首古風版,巨狐那一鏡之後留一段空白,再接金光覺醒的段落。最後用 Whisper 單獨轉配樂軌,確認沒有人聲和歌詞。
前段配樂:ACE-Step 的 caption
原文:參數 bpm 128、key_scale E Minor、vocal_language unknown、thinking 開啟,lm_negative_prompt 填 vocals, singing, choir, humming 這類詞;用的是 seed 636617189 那一首
Tense ancient Chinese xianxia film score, instrumental only, for a demon fox appearing at a ruined shrine on a rainy night. Low drone and sub-bass, breathy xiao flute and low guqin plucks over a throbbing pulse, nervous pipa tremolo and string tremolo building, thundering Chinese war drums and booming hits at the peak, sudden near silence, then a shimmering guzheng glissando and a firm, determined heroic finish. Dark, cinematic, hybrid orchestral.前段配樂:lyrics 欄(只放段落標籤)
原文:純配樂的 lyrics 只寫段落標籤,標籤不要疊太多,否則可能被當成歌詞唱出來
[Intro - dark drone]
[Verse - xiao and pulse]
[Build - pipa tremolo]
[Climax - war drums]
[Breakdown]
[Outro - heroic resolve]覺醒段落:ACE-Step 的 caption
原文:參數同上;用了同一批的兩首,seed 2553289049 接在系統綁定之後,seed 3447676319 開頭那一記重擊放在劇名
Ancient Chinese xianxia film score, instrumental only: a magic golden seal awakens. A bright shimmering rising swell with a guzheng glissando and airy high strings, then a firm, determined heroic theme on low strings and deep Chinese war drums, ending on one strong final hit. Hybrid orchestral, modern trailer sound design, cinematic.覺醒段落:lyrics 欄
[Intro - shimmering swell]
[Chorus - heroic resolve]
[Outro - final hit]她起跳那一格在成品的 40.417 秒(第 12 鏡第 46 格)。覺醒段落裡最大的一記重擊,在它前面把同一小節(4 拍)多重複一次,往後移到 40.419 秒,跟起跳差 0.002 秒,拍子沒有亂。系統提問那段底下也重複了一小節,讓她回答完的最後一個字之後 0.06 秒剛好一記重擊。
慢動作那 4 秒,配樂用低通濾波一路悶到 280 Hz,只剩超低頻長音和每秒一下的心跳聲。劇名出現前 0.36 秒抽到幾乎無聲,劇名出現那一格再一記重擊。起跳和劇名那兩下,另外疊了一層用程式合成的低頻重擊;心跳聲和系統提示的「叮」聲也是程式合成的。
交給 Claude Code 做配樂的提示詞
整理版:依這支配樂當時給 Claude Code 的工作說明整理,拿掉了本機路徑
幫這支預告片做配樂,先研究再生成。
1. 參考片:量它開頭幾分鐘的聲音。用 demucs 分出人聲和配樂,用 librosa 每 0.5 秒量一次音量,再用 PANNs 標出每一秒的聲音類別。整理出配樂在哪裡進、哪裡堆高、哪裡突然抽空,對白時配樂比人聲小幾 dB,每一條附秒數
2. 照我的分鏡時間軸,逐鏡寫出情緒、配樂在做什麼、緊張點落在哪幾秒,做成一條 0 到 1 的目標曲線
3. 用本機的 ACE-Step 生至少 3 個純配樂候選,BPM 和調性用參數指定,不要寫進 caption。每個候選量音量和起音強度,跟目標曲線畫在同一張圖上,挑最貼的一版
4. 用 Whisper 單獨轉配樂軌,確認沒有人聲或歌詞
5. 混音:對白時配樂自動壓低,人聲至少比配樂大 11.5 dB;最重的一記重擊對準【畫面事件】那一格,重擊前先抽空;整支響度 -14 LUFS「斬妖系統,綁定成功。是否接受第一個任務,斬殺一隻妖怪?」這句用 Gemini 的語音模型 gemini-3.1-flash-tts-preview 錄,聲音選 Leda,一次錄完整句。錄好再加一層中等的金屬感:梳狀回音、環形調變、短殘響和帶通,音高和速度都不動。效果加得太重,混進去之後「綁定」會聽不清楚。
原文:2026-10-24 送給 gemini-3.1-flash-tts-preview(聲音 Leda)的內容
Say in standard Mandarin Chinese, in the voice of a lively, cute little girl (活潑可愛的小女生) speaking fast (語速快): bright, bouncy and excited, like a cheerful game-system fairy announcing, with only tiny pauses at the punctuation, and ask the last part as a bright question: 斬妖系統,綁定成功。是否接受第一個任務,斬殺一隻妖怪?第六步:剪接、字幕和音量
12 鏡硬切串接,每一鏡從 1312x736 縮放裁切成 1280x720、24 fps。第 9 鏡的系統提問要 4.1 秒才講得完,這一鏡後半段用補幀放慢約 2 倍,從 73 幀拉長到 118 幀;第 10 鏡剪到 84 幀,讓後面起跳那一格對上配樂的重擊。劇名卡用金色行楷,在全黑之後浮現,停約 2.3 秒。
字幕用 Whisper 抓每一鏡實際開口和講完的時間。字幕和系統字卡先用 Python 的 PIL 畫成透明圖,再用 ffmpeg 疊上去,因為這台的 ffmpeg 沒有 drawtext 濾鏡。最後整支做 loudnorm,量出來是 -14.0 LUFS、真峰值 -1.9 dBTP。
檢查對白有沒有聽清楚,也是用 Whisper 轉回文字。Whisper 一次聽 30 秒,句子剛好跨在 30 秒的邊界上會漏字,所以分成幾段轉,每段都短於 30 秒。
交給 Claude Code 剪接和上字幕的提示詞
整理版:依這支的剪接程式整理
把 clips 資料夾裡的 12 鏡照順序硬切接起來,輸出 1280x720、24 fps:
1. 每一鏡先縮放再裁成 1280x720,聲音轉成 48 kHz 立體聲
2. 有台詞的鏡頭,用 Whisper 轉出那一鏡實際開口和講完的秒數,字幕照這個時間出現,講完多留 0.3 秒
3. 字幕白字黑邊、置中放在下方;系統字卡金框、深色底、金色字,放在上方
4. 整支用 loudnorm 調到 -14 LUFS,真峰值不超過 -1.5 dBTP
5. 做完用 Whisper 分段轉一次(每段短於 30 秒),列出跟台詞對不上的字本機做法的限制
- 解析度:本機輸出 1312x736,成品裁成 1280x720;上面幾家雲端旗艦有 1080p 到 4K
- 速度:成片 1 秒約 4 分鐘,照這個速度,一集 2 分鐘約要 8 小時
- 電費沒有實測,跟雲端的價格比較沒有算電費
- 雲端價格是牌價,沒算重生的次數,也沒算折扣
- 角色圖和系統提示音還是用雲端服務(ChatGPT、Gemini)
- 鏡與鏡之間的雨聲是硬切,每一鏡生出來的雨聲音色不太一樣
- 第 11、12 鏡她的小腿是綁腿和布鞋,設定圖是赤腳穿草鞋
生成和剪接卡關時的處理
ChatGPT 沒收到附件,生出別的角度?
第一次附上第 11 鏡的最後一格時沒有傳成功,ChatGPT 生出一張側面的構圖。重新附圖再送一次,提示詞開頭加一句「請完全以這則訊息的附件為準(上一張側面角度的不算)」,第二張就照附件的背影和機位畫。
首尾幀接上去,開頭像頓了一下?
H3 是從靜止的第一格起步,生出來的前 3 格幾乎不動。逐格比對相鄰兩格的差異,從第 3 格開始接,接縫那一步的畫面變化就跟前一鏡跑步時差不多。
系統音的「綁定」被聽成別的字?
人聲單獨轉文字是對的,混進雨聲和配樂之後,Whisper 轉成「放定」。問題在「統」和「綁」之間那個小停頓:「綁」這種音要靠前面一小段安靜才聽得出來,雨聲和配樂把空檔填滿了。只在那 0.3 秒把雨聲和配樂壓低 10 dB 就轉對了;整句調大聲、整段壓低配樂都沒有效果。
ACE-Step 一啟動就開始下載模型?
macOS 的啟動腳本沒有讀專案裡的 .env,會自己去下載預設模型(約 8.5 GB)。改用環境變數指定 ACESTEP_CONFIG_PATH=acestep-v15-xl-turbo、ACESTEP_LM_MODEL_PATH=acestep-5Hz-lm-4B 再啟動,加上 HF_HUB_OFFLINE=1,就不會連網下載。
用到的工具
MiniMax H3,MiniMax 的影片模型,畫面和聲音一起生成;官方也有雲端 API。
ComfyUI,在本機跑影片模型的工具,H3 用它內建的範本;GGUF 模型檔要另外裝 ComfyUI-GGUF,量化檔來自 unsloth。
ACE-Step 1.5,開源的音樂生成模型,這支的配樂在本機用它生。
ChatGPT,角色設定圖和每一鏡的第一格用網頁版的生圖功能。
Gemini 語音生成說明,系統提示音用的語音模型,含聲音清單和 API 用法(英文)。
Whisper,OpenAI 開源的語音轉文字模型,用來抓字幕時間、檢查對白。
demucs、librosa、PANNs,分離人聲、量音量曲線、標聲音類別,用來分析參考配樂。
ffmpeg,剪接、補幀、疊字幕和調響度。
Claude Code,Anthropic 的 AI 程式助手,這支的生成、剪接和混音程式由它寫。