EPISODES
MS-ECRAHG8BEPISODE VIDEO

新作パズル ぶよぶよ 到着編

結合版(6チャプター+ゆめみん4声目ミックス・26.12秒)

架空の落ちものパズル『ぶよぶよ』のゲームCM、約26秒。そば屋が家のインターホンを鳴らし、迎え入れられ、「新しいゲーム買ったよ!」と新作を披露する。福ちゃん・やめ太郎・たこさん・ゆめみんが「よし!やろうぜ!」と乗り、全員のアイコンが落ちものパズルの盤面で連鎖していく。紫がそば屋、赤がやめ太郎で、4つつながると消える。締めはナレーションの「(そば屋の腹は)ぶよぶよ!近日発売」。MiniMax H3で6チャプター生成し、H3のR2V音声上限3本を超えるゆめみんの鳴き声だけローカルでミックスした。

2026/08/271 VIDEO5 CASTv1 · MiniMax H3 (ComfyUI / Google Colab)
GENERATION v1

MiniMax H3 6チャプター生成(Colab)

MiniMax H3 (ComfyUI / Google Colab)結合版(6チャプター+ゆめみん4声目ミックス・26.12秒)
PROMPTrevision 1

MiniMax H3 チャプター別プロンプト(ch1〜ch6 逐語)

## ch1

The video starts EXACTLY on the attached first frame and ends EXACTLY on the attached last frame. Exactly one physical character is visible: Sobaya — the hulking 180cm/100kg muscular gray-skinned man whose face is always covered by a white mask with red markings, wearing a white T-shirt and carrying one standard glossy rectangular home-console game case and one large beer mug. Start on a WIDE eye-level exterior shot of a sunlit residential front door in BRIGHT MIDDAY DAYLIGHT, clear sky, DAYTIME NOT night. Sobaya walks toward the door and presses the intercom once; by the end the door remains CLOSED. The game case stays CLOSED in his hand and is an ordinary retail package with a blank non-readable cover, NOT a block-built box and NOT a block-like illustration. He keeps the ONE beer mug upright in his other hand and does not drink. Exactly one Sobaya appears at all times; he does not split or duplicate; no extra people, no duplicate case, no duplicate mug. No one speaks: no dialogue, no narration, no voice at all. do NOT render any on-screen text — no subtitles, no captions, no lettering, no Japanese characters; the video must contain no text at all. Cinematic live-action Japanese home comedy with photorealistic Sobaya and Fukuchan, matte 3D chibi Yametaro, and canonical stylized Takosan and Yumemin, all consistently lit in bright midday daylight; the falling-puzzle game uses glossy colorful round blob pieces with expressive face designs in a clean original TV game interface; no anime illustration look, no flat 2D cartoon. Soundscape: quiet residential street, light footsteps, one clear intercom chime, soft daytime ambience. The camera uses a slow tracking push-in with small amplitude at slow speed. Music: no background music.


## ch2

The video starts EXACTLY on the attached first frame and ends EXACTLY on the attached last frame. Start with the front door CLOSED in BRIGHT MIDDAY DAYLIGHT. Exactly two physical characters appear by the end, each exactly once: Sobaya, the hulking 180cm/100kg masked gray-skinned man whose face is always covered by a white mask with red markings, wearing a white T-shirt and holding one CLOSED ordinary rectangular game case and one upright beer mug, and Fukuchan, the slim stylish black-haired man in a loose black long coat over a white graphic T-shirt with a blank lanyard name tag. Fukuchan opens the CLOSED door once; Sobaya crosses the threshold as ONE continuous person, never duplicated. By the end the door is OPEN behind them, Sobaya still holds the CLOSED case and mug, and no other members have entered the frame. The case is ordinary retail packaging with no block illustration and no readable cover text. No one drinks. No one speaks: no dialogue, no narration, no voice at all. do NOT render any on-screen text — no subtitles, no captions, no lettering, no Japanese characters; the video must contain no text at all. Cinematic live-action Japanese home comedy with photorealistic Sobaya and Fukuchan, matte 3D chibi Yametaro, and canonical stylized Takosan and Yumemin, all consistently lit in bright midday daylight; the falling-puzzle game uses glossy colorful round blob pieces with expressive face designs in a clean original TV game interface; no anime illustration look, no flat 2D cartoon. Soundscape: one door latch click, soft hinges, two sets of footsteps, quiet bright home ambience. Use a locked-off static camera at eye level. Music: no background music.


## ch3

Required attached input files: <Picture 1> = clip3_start.png — start keyframe, the video's FIRST frame; <Picture 2> = clip3_end.png — end keyframe, the video's LAST frame; <Picture 3> = Sobaya_sheet.png — Sobaya's character model sheet, identity/design reference only, NOT a composition reference; <Picture 4> = Fukuchan_sheet.png — Fukuchan's character model sheet, identity/design reference only, NOT a composition reference; <Picture 5> = Yametaro_sheet.png — Yametaro's character model sheet, identity/design reference only, NOT a composition reference; <Picture 6> = Takosan_sheet.png — Takosan's character model sheet, identity/design reference only, NOT a composition reference; <Picture 7> = Yumemin_sheet.png — Yumemin's character model sheet, identity/design reference only, NOT a composition reference; <Picture 8> = height_lineup.png — relative height/scale reference only, NOT a composition reference; <Audio 1> = clip3_line1_sobaya_h3pad.wav — Sobaya's spoken line, use AS-IS as the dialogue audio. These attachments are REQUIRED inputs and must remain attached for this generation. The video starts EXACTLY on <Picture 1> and ends EXACTLY on <Picture 2>. In a bright living room in BRIGHT MIDDAY DAYLIGHT, exactly five physical characters are visible once each: Sobaya (<Picture 3>), Fukuchan (<Picture 4>), Yametaro (<Picture 5>), Takosan (<Picture 6>), and Yumemin (<Picture 7>). The TV is OFF and black. Sobaya moves the CLOSED ordinary rectangular game case from his hand to the table, opens it, and presents the single visible game media item; by the end the case is OPEN on the table and the media remains inside it. The package is normal glossy retail packaging with a blank non-readable cover, NOT a block-built box and NOT a block-like illustration. ONLY Sobaya (S1) speaks — he says <d>[Japanese] 新しいゲーム買ったよ!</d>, lip-syncing to <Audio 1>; he begins the line almost immediately and the mouth opening in his mask moves ONLY while the spoken line in <Audio 1> is sounding; the trailing silence at the end of <Audio 1> is padding, and through it and for the rest of the chapter his mouth stays CLOSED. Use <Audio 1> AS-IS as the dialogue audio and do NOT generate any voice — no synthesized speech, no narration, no second voice. Fukuchan (<Picture 4>), Yametaro (<Picture 5>), Takosan (<Picture 6>) and Yumemin (<Picture 7>) do NOT speak; their mouths and faces stay silent and reactive. Exactly one of each character appears; no duplicates, no duplicate case, no duplicate mug, no one drinks. Takosan keeps two human arms and six tentacles; Yumemin keeps the blue body, trunk-like nose, and white rear. The reference sheets' text labels must NOT appear in the video. do NOT render any on-screen text — no subtitles, no captions, no lettering, no Japanese characters; the video must contain no text at all. Cinematic live-action Japanese home comedy with photorealistic Sobaya and Fukuchan, matte 3D chibi Yametaro, and canonical stylized Takosan and Yumemin, all consistently lit in bright midday daylight; the falling-puzzle game uses glossy colorful round blob pieces with expressive face designs in a clean original TV game interface; no anime illustration look, no flat 2D cartoon. Soundscape: soft case hinge, one gentle media click, bright living-room ambience, quiet excited reactions with no intelligible words. Use a locked-off static camera at a slight low angle. Music: no background music.


## ch4

Required attached input files: <Picture 1> = clip4_start.png — start keyframe, the video's FIRST frame; <Picture 2> = clip4_end.png — end keyframe, the video's LAST frame; <Picture 3> = Sobaya_sheet.png — Sobaya's character model sheet, identity/design reference only, NOT a composition reference; <Picture 4> = Fukuchan_sheet.png — Fukuchan's character model sheet, identity/design reference only, NOT a composition reference; <Picture 5> = Yametaro_sheet.png — Yametaro's character model sheet, identity/design reference only, NOT a composition reference; <Picture 6> = Takosan_sheet.png — Takosan's character model sheet, identity/design reference only, NOT a composition reference; <Picture 7> = Yumemin_sheet.png — Yumemin's character model sheet, identity/design reference only, NOT a composition reference; <Picture 8> = height_lineup.png — relative height/scale reference only, NOT a composition reference; <Audio 1> = clip4_line1_fukuchan_h3pad.wav — Fukuchan's spoken line, use AS-IS as the dialogue audio; <Audio 2> = clip4_line2_yametaro.wav — Yametaro's spoken line, use AS-IS as the dialogue audio; <Audio 3> = clip4_line3_takosan_h3pad.wav — Takosan's spoken line, use AS-IS as the dialogue audio. These attachments are REQUIRED inputs and must remain attached for this generation. The video starts EXACTLY on <Picture 1> and ends EXACTLY on <Picture 2>, in the same bright living room in BRIGHT MIDDAY DAYLIGHT. Exactly five physical characters appear once each: Sobaya (<Picture 3>), Fukuchan (<Picture 4>), Yametaro (<Picture 5>), Takosan (<Picture 6>), and Yumemin (<Picture 7>). The OPEN ordinary game case and its ONE media item remain on the table. The TV powers ON and shows a clean empty ready board of glossy round face blobs, not a block illustration. Fukuchan (S1), Yametaro (S2) and Takosan (S3) all say the same Japanese line <d>[Japanese] よし!やろうぜ!</d> at the SAME time, each lip-syncing to his own track: Fukuchan to <Audio 1>, Yametaro to <Audio 2>, Takosan to <Audio 3>. The three lines start together almost immediately and intentionally overlap. Use <Audio 1>, <Audio 2> and <Audio 3> AS-IS as the dialogue audio and do NOT generate any voice — no synthesized speech, no narration, no extra voice. The trailing silence at the end of <Audio 1> and <Audio 3> is padding; through it and for the rest of the chapter those two mouths stay CLOSED. Yumemin (<Picture 7>) opens her mouth in one short simultaneous cheer at the same moment and closes it right after, but generate NO voice and NO sound for her. Sobaya (<Picture 3>) does NOT speak; the mouth opening in his mask stays CLOSED for the whole chapter. Takosan keeps two human arms and six tentacles; Yumemin keeps the blue body, trunk-like nose, and white rear. Exactly one of each character, no duplicates, no extra props, no readable package text, no one drinks. The reference sheets' text labels must NOT appear in the video. do NOT render any on-screen text — no subtitles, no captions, no lettering, no Japanese characters; the video must contain no text at all. Cinematic live-action Japanese home comedy with photorealistic Sobaya and Fukuchan, matte 3D chibi Yametaro, and canonical stylized Takosan and Yumemin, all consistently lit in bright midday daylight; the falling-puzzle game uses glossy colorful round blob pieces with expressive face designs in a clean original TV game interface; no anime illustration look, no flat 2D cartoon. Soundscape: TV power-on chime, one controller click, the three intentionally overlapping supplied voices, and a bright game-start sparkle. The camera makes a slow push-in with small amplitude at slow speed. Music: no background music.


## ch5

The video starts EXACTLY on the attached first frame and ends EXACTLY on the attached last frame. Start on a CLOSE-UP straight-on television screen in BRIGHT MIDDAY DAYLIGHT with broad soft reflections on the glass, NOT night and NOT a dark neon room. Show an original clean falling-puzzle game interface: a vertical grid, glossy round blob pieces, readable color groups, pieces dropping from above, and satisfying four-match clearing. The five canonical designs appear as game blobs: Sobaya's white mask with red marks, Fukuchan's black-haired smiling face, Yametaro's bowl hair and round glasses, Takosan's black hood and flat white face, and Yumemin's blue trunk-nose mascot. Exactly zero physical characters and zero duplicate TV screens are visible. Follow the falling-and-chain sequence anchored by the first and last frames, preserving its order, timing and reaction structure; do not invent a different chain. PURPLE-group pieces are Sobaya (white mask with red markings) and RED-group pieces are Yametaro (black bowl hair and round glasses). Four connected same-color pieces disappear according to the game's rule, then the next chain continues. Repeated blobs are intentional game elements only. No one speaks: no dialogue, no narration, no voice at all; only distant non-verbal laughter. do NOT render any on-screen text — no subtitles, no captions, no lettering, no Japanese characters; no brand logos, no copied game assets, no UI labels. Cinematic live-action Japanese home comedy with photorealistic Sobaya and Fukuchan, matte 3D chibi Yametaro, and canonical stylized Takosan and Yumemin, all consistently lit in bright midday daylight; the falling-puzzle game uses glossy colorful round blob pieces with expressive face designs in a clean original TV game interface; no anime illustration look, no flat 2D cartoon. Soundscape: soft falling-piece ticks, four-match chime, bright clear pop, rising chain jingle, second four-piece flash, and distant non-verbal laughter. The camera makes a slow push-in with small amplitude at slow speed. Music: no background music.


## ch6

Required attached input files: <Picture 1> = clip6_start.png — start keyframe, the video's FIRST frame; <Picture 2> = clip6_end.png — end keyframe, the video's LAST frame; <Picture 3> = Sobaya_sheet.png — Sobaya's character model sheet, identity/design reference only, NOT a composition reference; <Picture 4> = Takosan_sheet.png — Takosan's character model sheet, identity/design reference only, NOT a composition reference; <Picture 5> = Fukuchan_sheet.png — Fukuchan's character model sheet, identity/design reference only, NOT a composition reference; <Picture 6> = Yametaro_sheet.png — Yametaro's character model sheet, identity/design reference only, NOT a composition reference; <Picture 7> = Yumemin_sheet.png — Yumemin's character model sheet, identity/design reference only, NOT a composition reference; <Audio 1> = clip6_line1_narrator.wav — the off-screen Narrator's line, use AS-IS as the narration audio. These attachments are REQUIRED inputs and must remain attached for this generation. The video starts EXACTLY on <Picture 1> and ends EXACTLY on <Picture 2>. This is a deliberate CUT from the game screen to a CLOSE-UP original colorful game poster at a slight low angle, and the camera pushes in with medium amplitude at slow speed. Use bright midday-like clean studio lighting, not night. The artwork contains one large Sobaya mask game piece (<Picture 3>) plus exactly one Fukuchan face piece (<Picture 5>), one Yametaro face piece (<Picture 6>), one Takosan face piece (<Picture 4>), and one Yumemin face piece (<Picture 7>). They are glossy round game pieces only, not physical bodies, and no mouth on the poster moves. ONLY an off-screen Narrator (S1) is heard, saying <d>[Japanese] (そば屋の腹は)ぶよぶよ!近日発売</d>. Use <Audio 1> AS-IS as the narration audio and do NOT generate any voice — no synthesized speech, no character voice, no second voice. Do NOT render lettering in the generated poster; the Japanese overlay and the VOICEVOX credit are added later in CapCut. do NOT render any on-screen text — no subtitles, no captions, no lettering, no Japanese characters, no logos, no watermark; the video must contain no text at all. Cinematic live-action Japanese home comedy with photorealistic Sobaya and Fukuchan, matte 3D chibi Yametaro, and canonical stylized Takosan and Yumemin, all consistently lit in bright midday daylight; the falling-puzzle game uses glossy colorful round blob pieces with expressive face designs in a clean original TV game interface; no anime illustration look, no flat 2D cartoon. Soundscape: short chain sparkle tail, clean poster reveal whoosh, and the supplied narration. Music: no background music.
INPUTS24
AUDIO

ch3 そば屋「新しいゲーム買ったよ!」(H3投入・パディング済み)

H3投入音声 · Audio 1 (ch3) · clip3_line1_sobaya_h3pad.wav

1.0x候補を採用。正典のモンスターボイス処理込み。H3の最短2.0秒制約に合わせて末尾をパディングした別名ファイル。

AUDIO

ch4 福ちゃん「よし!やろうぜ!」(H3投入・パディング済み)

H3投入音声 · Audio 1 (ch4) · clip4_line1_fukuchan_h3pad.wav

1.5x候補を採用。3人同時発話の1本目。

AUDIO

ch4 やめ太郎「よし!やろうぜ!」(H3投入)

H3投入音声 · Audio 2 (ch4) · clip4_line2_yametaro.wav

1.5x候補を採用。2.0秒以上あるためパディング不要。3人同時発話の2本目。

AUDIO

ch4 たこさん「よし!やろうぜ!」(H3投入・パディング済み)

H3投入音声 · Audio 3 (ch4) · clip4_line3_takosan_h3pad.wav

1.0x候補を採用。3人同時発話の3本目で、H3のR2V音声スロットはここで上限3本に達する。

AUDIO

ch6 ナレーション「(そば屋の腹は)ぶよぶよ!近日発売」(H3投入)

H3投入音声 · Audio 1 (ch6) · clip6_line1_narrator.wav

4.20秒の1.0xテイクをそのまま採用(前ランからの流用)。画面外ナレーションだがH3にそのまま鳴らさせている。

IMAGE

そば屋 キャラクターモデルシート

キャラクター参照 · Sobaya_sheet.png
IMAGE

福ちゃん キャラクターモデルシート

キャラクター参照 · Fukuchan_sheet.png
IMAGE

無職やめ太郎 キャラクターモデルシート

キャラクター参照 · Yametaro_sheet.png
IMAGE

たこさん キャラクターモデルシート

キャラクター参照 · Takosan_sheet.png

2本の人間の腕と6本の触手を維持することが必須条件。

IMAGE

ゆめみん キャラクターモデルシート

キャラクター参照 · Yumemin_sheet.png

青い体・象の鼻のような突起・白いお尻を維持することが必須条件。

IMAGE

身長比較(スケール参照)

キャラクター参照 · height_lineup.png

相対的な身長・スケール関係のみの参照。構図・レイアウト・ラベル・背景は引き継がせない。

IMAGE

ch1 開始キーフレーム

キーフレーム · clip1_start.png
IMAGE

ch1 終了キーフレーム(=ch2開始)

キーフレーム · clip1_end.png
IMAGE

ch2 開始キーフレーム(境界共有)

キーフレーム · clip2_start.png
IMAGE

ch2 終了キーフレーム

キーフレーム · clip2_end.png
IMAGE

ch3 開始キーフレーム

キーフレーム · Picture 1 · clip3_start.png
IMAGE

ch3 終了キーフレーム(=ch4開始)

キーフレーム · Picture 2 · clip3_end.png
IMAGE

ch4 開始キーフレーム(境界共有)

キーフレーム · Picture 1 · clip4_start.png
IMAGE

ch4 終了キーフレーム

キーフレーム · Picture 2 · clip4_end.png
IMAGE

ch5 開始キーフレーム(ゲーム画面)

キーフレーム · clip5_start.png
IMAGE

ch5 終了キーフレーム(連鎖後の盤面)

キーフレーム · clip5_end.png
IMAGE

ch6 開始キーフレーム(ゲームポスター)

キーフレーム · Picture 1 · clip6_start.png
IMAGE

ch6 終了キーフレーム

キーフレーム · Picture 2 · clip6_end.png
AUDIO

ch4 ゆめみん「きゅー!」(ローカル合成・4声目)

ローカル合成音声 · clip4_line4_yumemin.wav

0.49秒の非言語の鳴き声。H3のR2Vは音声3本までのため投入できず、結合時にffmpegのamixで他3声と同じt=0地点へ重ねた。H3側にはゆめみんの口を開閉させつつ音は生成させない指示を入れている。