跳至主要內容

SHOWCASE · 15 FILMS · 43 RESULTS

這一頁的每一個畫面,都在作者家裡生成。

這裡的圖片與影片,全部由 Phasyn Studio 透過控制器 API(t2i‑pc、fl2va‑pc、ref2va‑pc)在作者的兩台機器、三顆 GPU 上生成。每個單次生成的結果都附上 Prompt、管線、每一步跑在哪台機器,以及實測時間;短片則列出它剪進的每一個鏡頭。

支短片,共 5.4 分鐘
15
個生成結果
43
端到端生成時間合計
67.9分鐘
顆 GPU,分在 2 台機器
3
AI 生成 · Powered by MiniMax H3

屋頂信使短片ref2va-pcfl2va-pc

15 個鏡頭1920×1080配樂:Suno

58.4 秒

精選

看全部結果

一段 Prompt,一張圖,一段有聲影片。

雨夜的巷子:t2i‑pc 先生成 1152×672 的圖,縮小成 768×448 後當成 fl2va‑pc 的第一幀,再加上動作與聲音的描述,生成 5.17 秒的有聲影片。

含 AI 生成內容(MiniMax H3、Qwen-Image 2.1)

幕後文字 → 圖片 → 影片

三種風格,同一個流程:每一段都是原生靜圖慢慢推近,再溶接到由它長出來的 FL2VA 影片,聲音隨之淡入。

24.8 秒 · 6 個鏡頭

t2i-pcfl2va-pc配樂:Suno

原生影片放大 1.75 倍、疊在模糊背景上(只重新取樣,無超解析)

看鏡頭清單
鏡頭執行種子長度
1577fc7a4474.96 s
27d52714a4044.96 s
356bca2e851112.50 s
43553e6f051114.96 s
5a291923f52122.46 s
6964fa97152124.96 s
  1. 01圖片

    雨夜的巷子
    AI 生成 · Qwen-Image 2.1

    t2i‑pc · Qwen‑Image 2.1 · 種子 23 · 端到端 26.9 秒

    圖片的 Prompt

    Cinematic medium-wide shot, high-end anime film still. At night a young woman with long straight silver-white hair with faint cyan highlights, calm violet-grey eyes, wearing a dark navy hooded travel cloak edged with a thin luminous cyan-to-violet trim, a small hair ornament of three stacked translucent glass frames walks along a rain-soaked narrow street lined with glowing shop signs in cyan, blue and violet, holding a translucent umbrella; rain streaks catch the light, puddles mirror the neon, soft bokeh in the distance, gentle steam rising from a vent. Moody, calm, detailed, no readable text, no logo, no watermark.

  2. 02有聲影片

    AI 生成 · Powered by MiniMax H3

    fl2va‑pc · MiniMax H3 · 種子 202 · 124 幀 · 端到端 57.2 秒

    影片的 Prompt(動作與聲音)

    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] High-end anime film style, night, cool neon light. The young woman with silver-white hair under the translucent umbrella walks slowly forward along the rain-soaked narrow street toward the camera. Heavy rain falls in fine bright streaks through the cyan and violet glow of the shop signs, droplets run along the edge of her umbrella and splash in the puddles that mirror the neon. Steam drifts from a vent on the right. The camera tracks backward in front of her at walking pace, keeping her centered, with soft bokeh lights shifting in the background. She glances up from under the umbrella toward the sky. overall_soundscape: Dense steady rain on the umbrella and pavement, water dripping from awnings, soft footsteps through shallow puddles, a faint hum of neon signs and a distant city murmur. non_diegetic_music: N/A

這 57.2 秒,每台機器在做什麼

  1. 文字編碼DGX Spark · GB101.3 秒 @ 0.0s
  2. 影片 VAE 編碼RTX 50801.1 秒 @ 1.4s
  3. DiT 階段RTX 508032 秒 @ 2.5s
  4. 影片 VAE 解碼RTX 508022.4 秒 @ 34.4s
  5. 音訊 VAE 解碼RTX 30600.9 秒 @ 34.4s
  6. 封裝(Mux)CPU0.3 秒 @ 56.8s
步驟的位置依照這次執行在 Records 裡的 Step Timeline(下方是實際畫面)。影片解碼在 RTX 5080、音訊解碼在 RTX 3060,兩者同時進行。
這次執行在 Records 裡的紀錄:每一步、每台機器(實際畫面)
這次執行在 Records 裡的紀錄:每一步、每台機器(實際畫面)端到端 57.2 秒

同一個角色,換一個場景。

先用 t2i‑pc 生成白底的角色設定圖,再把它當成 ref2va‑pc 唯一的參考圖。Prompt 用 <Subject 1> 指名她,並要求 H3 保留 <Picture 1> 裡的樣子。請自己比較每一段。

角色設定圖:The CourierAI 生成 · Qwen-Image 2.1
參考圖 · 768×1024t2i-pc · Qwen-Image 2.1 · 端到端 26 秒

三個候選種子,各 26–37 秒

  • 種子 77
  • 種子 4242
  • 種子 20262026選用

同一張參考圖,11 個場景

每一段都只用這張圖當參考;場景、動作與鏡頭全由 Prompt 決定。

  • AI 生成 · Powered by MiniMax H310.13 s · 243 f

    參考圖 → 影片

    屋頂信使・暮色啟程

    ref2va-pc
    端到端
    195 秒
    解析度
    768×448
    種子
    9011
    • DGX Spark · GB108.5 秒文字編碼
    • RTX 5080186 秒影片編碼 9.2DiT 129影片解碼 47.8
    • RTX 30606.5 秒音訊編碼 1.2音訊解碼 5.3
    • CPU0.4 秒封裝

    用在屋頂信使上線預告・信使

    看完整 Prompt

    subject_definitions: <Subject 1> is the young woman shown in <Picture 1>, with long silver-white hair with faint cyan highlights, calm violet-grey eyes, a dark navy hooded travel cloak edged with a thin luminous cyan-to-violet trim, a small hair ornament of three stacked translucent glass frames, brown leather boots and a leather satchel on a cross-body strap. summary: [reference generation] At dusk <Subject 1> stands at the edge of a high rooftop above a vast hillside city, looks at a small glowing parcel and tucks it into her satchel, keeping the exact appearance from <Picture 1>. retention_analysis: <Subject 1> (appears in [Shot 1]): fully_preserved - face, silver-white hair with cyan highlights, violet-grey eyes, the navy cloak with its luminous cyan-to-violet trim, the glass-frame hair ornament, boots and satchel. detailed_description: High-end 2D anime film style, dusk light fading from warm orange to deep blue, rich painterly backgrounds and crisp clean lineart. [Shot 1] A wide shot from behind and slightly above. <Subject 1> stands at the low parapet of a high concrete rooftop, her cloak and long silver-white hair moving in the evening wind, a vast hillside city spreading below toward the sea. She lifts a small parcel wrapped in pale paper and tied with string; it glows softly cyan from inside, lighting her fingers and face. She looks at it for a moment, then tucks it carefully into her leather satchel and closes the flap. Across the city, windows and street lights switch on one district after another. The camera pushes in slowly and arcs around to her profile, ending on her calm, determined face against the darkening sky. overall_soundscape: Steady evening wind across the rooftop, the flutter of the cloak, a faint crystalline chime from the glowing parcel, distant traffic and a far-off train horn, birds settling for the night. non_diegetic_music: N/A

    MiniMax H3dit5080-ref2va-fp8-pai8 (H3 Ref2VA · FP8 · PAI 8 steps)執行 c11f2c4b2026-10-18

  • AI 生成 · Powered by MiniMax H310.13 s · 243 f

    參考圖 → 影片

    屋頂信使・雨夜疾奔

    ref2va-pc
    端到端
    180 秒
    解析度
    768×448
    種子
    9031
    • DGX Spark · GB108.4 秒文字編碼
    • RTX 5080171 秒影片編碼 5.1DiT 123影片解碼 43.5
    • RTX 30602.1 秒音訊編碼 1.0音訊解碼 1.1
    • CPU0.4 秒封裝

    用在上線預告・信使

    看完整 Prompt

    subject_definitions: <Subject 1> is the young woman shown in <Picture 1>, with long silver-white hair with faint cyan highlights, calm violet-grey eyes, a dark navy hooded travel cloak edged with a thin luminous cyan-to-violet trim, a small hair ornament of three stacked translucent glass frames, brown leather boots and a leather satchel on a cross-body strap. summary: [reference generation] <Subject 1> sprints across rain-soaked rooftops at night, splashing through neon puddles and vaulting a low pipe, keeping the exact appearance from <Picture 1>. retention_analysis: <Subject 1> (appears in [Shot 1]): fully_preserved - face, silver-white hair with cyan highlights, violet-grey eyes, the navy cloak with its luminous cyan-to-violet trim, the glass-frame hair ornament, boots and satchel. detailed_description: High-end 2D anime film style, night, heavy rain, cyan and violet neon reflections, crisp clean lineart. [Shot 1] A side tracking shot at rooftop height. <Subject 1>, hood up, sprints from left to right across a chain of flat wet rooftops. Her boots splash through shallow puddles that mirror glowing shop signs below; spray flies with each step. She vaults cleanly over a low horizontal pipe with one hand and keeps running without losing speed. Her cloak streams behind her and its cyan-to-violet trim leaves a faint glowing trail in the rain. Water tanks, antennas and dripping laundry lines rush past in the foreground. The camera tracks alongside her at the same speed, slightly ahead, with rain streaks catching the neon light. overall_soundscape: Heavy steady rain, fast splashing footsteps on wet concrete, rushing wind, the snap of the cloak, a distant roll of thunder. non_diegetic_music: N/A

    MiniMax H3dit5080-ref2va-fp8-pai8 (H3 Ref2VA · FP8 · PAI 8 steps)執行 a5fcea352026-10-18

  • AI 生成 · Powered by MiniMax H35.17 s · 124 f

    參考圖 → 影片

    屋頂飛奔

    ref2va-pc
    端到端
    108 秒
    解析度
    768×448
    種子
    505
    • DGX Spark · GB108.5 秒文字編碼
    • RTX 508099.2 秒影片編碼 5.7DiT 70.7影片解碼 22.8
    • RTX 30602.7 秒音訊編碼 1.7音訊解碼 1.0
    • CPU0.5 秒封裝
    看完整 Prompt

    subject_definitions: <Subject 1> is the young woman shown in <Picture 1>, with long silver-white hair with faint cyan highlights, calm violet-grey eyes, a dark navy hooded travel cloak edged with a thin luminous cyan-to-violet trim, a small hair ornament of three stacked translucent glass frames, brown leather boots and a leather satchel on a cross-body strap. summary: [reference generation] <Subject 1> sprints across city rooftops at sunset and leaps over a wide gap between two buildings, keeping the exact appearance from <Picture 1>. retention_analysis: <Subject 1> (appears in [Shot 1]): fully_preserved - face, silver-white hair with cyan highlights, violet-grey eyes, the navy cloak with its luminous cyan-to-violet trim, the glass-frame hair ornament, boots and satchel. detailed_description: High-end 2D anime film style with warm sunset light, rich painterly backgrounds and crisp clean lineart. [Shot 1] A side tracking shot at rooftop height over a dense hillside city at sunset. <Subject 1> runs fast from left to right across flat concrete rooftops, her long silver-white hair and her cloak streaming behind her, the cyan-to-violet trim of the cloak leaving a faint glowing trail. Water tanks, antennas and lines of laundry rush past in the foreground and blur with speed. She reaches the edge of the roof, plants one boot on the low parapet and leaps across the wide gap between two buildings. For a moment she is airborne against a sky that fades from orange to violet, with the first city lights glowing far below and birds scattering from a wire. She lands smoothly on the next roof in a low crouch, one hand touching the concrete, and immediately rises and keeps running toward the right edge of the frame. The camera tracks alongside her at the same speed, slightly ahead of her, rises with the jump and dips a little as she lands. Warm backlight outlines her silhouette and catches loose strands of hair. overall_soundscape: Strong rushing wind, quick light footsteps on concrete, the snap and flutter of the cloak, a soft thud on landing, birds taking off, distant city traffic far below. non_diegetic_music: N/A

    MiniMax H3dit5080-ref2va-fp8-pai8執行 f083d1662026-10-18

  • AI 生成 · Powered by MiniMax H35.17 s · 124 f

    參考圖 → 影片

    山丘上的草原

    ref2va-pc
    端到端
    105 秒
    解析度
    768×448
    種子
    606
    • DGX Spark · GB108.3 秒文字編碼
    • RTX 508096.5 秒影片編碼 5.5DiT 68.5影片解碼 22.5
    • RTX 30602.4 秒音訊編碼 1.5音訊解碼 0.9
    • CPU0.3 秒封裝
    看完整 Prompt

    subject_definitions: <Subject 1> is the young woman shown in <Picture 1>, with long silver-white hair with faint cyan highlights, calm violet-grey eyes, a dark navy hooded travel cloak edged with a thin luminous cyan-to-violet trim, a small hair ornament of three stacked translucent glass frames, brown leather boots and a leather satchel on a cross-body strap. summary: [reference generation] On a bright windy hilltop meadow, <Subject 1> walks through tall silver grass and catches a small floating glass frame that flashes in the sunlight, keeping the exact appearance from <Picture 1>. retention_analysis: <Subject 1> (appears in [Shot 1]): fully_preserved - face, silver-white hair with cyan highlights, violet-grey eyes, the navy cloak with its luminous cyan-to-violet trim, the glass-frame hair ornament, boots and satchel. detailed_description: High-end 2D anime film style, bright late-morning daylight, saturated blue sky with towering white clouds, crisp clean lineart and soft painterly shading. [Shot 1] A slow orbiting medium-wide shot on a grassy hilltop. <Subject 1> walks forward through waist-high silver grass dotted with small blue flowers while a strong breeze sends waves rolling across the meadow and lifts her silver-white hair and the hem of her navy cloak. A small translucent glass frame drifts down through the air in front of her, turning slowly and flashing with cyan and violet reflections of the sun. She stops, raises her right hand and catches it gently between two fingers, then lifts it up to look through it toward the sky with a calm, curious smile. Grass seeds and a few petals float through the sunlight around her. The camera orbits slowly from her left side to the front of her, keeping her in the center of the frame against the clouds. overall_soundscape: A steady warm breeze over the hilltop, grass rustling in waves, distant birdsong, a soft crystalline ring when the glass frame touches her fingers. non_diegetic_music: N/A

    MiniMax H3dit5080-ref2va-fp8-pai8執行 5adf42372026-10-18

  • AI 生成 · Powered by MiniMax H35.17 s · 124 f

    參考圖 → 影片

    屋頂信使・躍過樓隙

    ref2va-pc
    端到端
    103 秒
    解析度
    768×448
    種子
    9041
    • DGX Spark · GB108.3 秒文字編碼
    • RTX 508094.5 秒影片編碼 5.0DiT 67.1影片解碼 22.4
    • RTX 30602.0 秒音訊編碼 1.0音訊解碼 1.0
    • CPU0.3 秒封裝

    用在上線預告・信使

    看完整 Prompt

    subject_definitions: <Subject 1> is the young woman shown in <Picture 1>, with long silver-white hair with faint cyan highlights, calm violet-grey eyes, a dark navy hooded travel cloak edged with a thin luminous cyan-to-violet trim, a small hair ornament of three stacked translucent glass frames, brown leather boots and a leather satchel on a cross-body strap. summary: [reference generation] <Subject 1> leaps across a wide gap between two buildings in the rain at night and lands in a crouch on the far roof, keeping the exact appearance from <Picture 1>. retention_analysis: <Subject 1> (appears in [Shot 1]): fully_preserved - face, silver-white hair with cyan highlights, violet-grey eyes, the navy cloak with its luminous cyan-to-violet trim, the glass-frame hair ornament, boots and satchel. detailed_description: High-end 2D anime film style, night rain, neon skyline, crisp clean lineart, dramatic low angle. [Shot 1] A low-angle shot looking up from inside the gap between two tall buildings. <Subject 1> runs to the edge of the left roof, plants one boot on the parapet and leaps across the wide gap. For a moment she is airborne against a rain-streaked skyline of glowing cyan and violet signs, her cloak spread wide and her hair flowing, rain drops frozen in the light. She lands smoothly on the far roof in a low crouch, one hand touching the wet concrete, water splashing around her. The camera tilts up to follow the jump and settles as she lands. overall_soundscape: Rushing wind and rain, a brief hush while she is in the air, a solid thud and splash on landing, distant thunder. non_diegetic_music: N/A

    MiniMax H3dit5080-ref2va-fp8-pai8 (H3 Ref2VA · FP8 · PAI 8 steps)執行 fd5913b22026-10-18

  • AI 生成 · Powered by MiniMax H35.17 s · 124 f

    參考圖 → 影片

    穿過那道框 1・接住

    ref2va-pc
    端到端
    175 秒
    解析度
    768×448
    種子
    606
    • DGX Spark · GB1016.1 秒文字編碼
    • RTX 5080158 秒影片編碼 11.4DiT 124影片解碼 22.7
    • RTX 30603.7 秒音訊編碼 2.4音訊解碼 1.3
    • CPU0.3 秒封裝

    參考圖角色設定圖:The Courier+ 臉部特寫(左右鏡像)

    看參考圖
    臉部特寫(左右鏡像)AI 生成 · Qwen-Image 2.1
    臉部特寫(左右鏡像) · t2i‑pc · 執行 ff115974 · 種子 2026
    看完整 Prompt

    subject_definitions: <Subject 1> is the young woman shown in <Picture 1> (full-body reference) and <Picture 2> (head-and-shoulders close-up of the same woman), with long silver-white hair with faint cyan highlights, violet-grey eyes (soft lavender-grey irises), a dark navy hooded travel cloak edged with a thin luminous cyan-to-violet trim, a small hair clip of three overlapping flat translucent glass rectangles with thin glowing cyan edges worn on the right side of her head (viewer's left, exactly as in <Picture 1>), brown leather boots and a leather satchel on a cross-body strap. <Picture 1> provides her outfit, proportions, boots, satchel and the position of her hair clip; <Picture 2> provides her exact face and eyes. summary: [reference generation] On a windy hilltop meadow in golden late-afternoon light, <Subject 1> walks through tall silver grass and catches a small floating glass frame that flashes in the sun, keeping the exact appearance from <Picture 1> and the exact face from <Picture 2>. retention_analysis: <Subject 1> (appears in [Shot 1]): fully_preserved - face and violet-grey eyes from <Picture 2>; silver-white hair with cyan highlights, the glass hair clip on the right side of her head, the navy cloak with its luminous cyan-to-violet trim, boots and satchel from <Picture 1>. detailed_description: High-end anime film style matching cinematic key art: painterly backgrounds, soft rim light, gentle volumetric sunlight from a low warm sun, crisp clean lineart and rich soft shading, not flat. [Shot 1] A slow orbiting medium shot on a grassy hilltop, <Subject 1> framed from head to knees with her head always fully in frame, hood down. She walks forward through waist-high silver grass while a steady breeze sends waves across the meadow and lifts her hair and the hem of her navy cloak. A small flat rectangular glass frame, like an empty picture frame, drifts down in front of her, turning slowly and flashing cyan and violet. At about two seconds she stops, raises her right hand and catches it gently between two fingers, then lifts it in front of one eye and looks through it toward the sky with a calm, curious smile. The camera orbits slowly from her left side to the front, keeping her centered against warm clouds. overall_soundscape: A steady warm breeze over the hilltop, grass rustling in waves, distant birdsong, and one clear bright crystalline ring, close and prominent, at the instant the glass frame touches her fingers. non_diegetic_music: N/A

    MiniMax H3dit5080-ref2va-fp8-pai8 (H3 Ref2VA · FP8 · PAI 8 steps)執行 7db436fc2026-10-19

  • AI 生成 · Powered by MiniMax H35.17 s · 124 f

    參考圖 → 影片

    穿過那道框 2・框裡的天空

    ref2va-pc
    端到端
    168 秒
    解析度
    768×448
    種子
    6621
    • DGX Spark · GB1016 秒文字編碼
    • RTX 5080151 秒影片編碼 10DiT 118影片解碼 22.7
    • RTX 30603.1 秒音訊編碼 2.2音訊解碼 0.9
    • CPU0.4 秒封裝

    參考圖角色設定圖:The Courier+ 臉部特寫(左右鏡像)

    看參考圖
    臉部特寫(左右鏡像)AI 生成 · Qwen-Image 2.1
    臉部特寫(左右鏡像) · t2i‑pc · 執行 ff115974 · 種子 2026
    看完整 Prompt

    subject_definitions: <Subject 1> is the young woman shown in <Picture 1> (full-body reference) and <Picture 2> (head-and-shoulders close-up of the same woman), with long silver-white hair with faint cyan highlights, violet-grey eyes (soft lavender-grey irises), a dark navy hooded travel cloak edged with a thin luminous cyan-to-violet trim, a small hair clip of three overlapping flat translucent glass rectangles with thin glowing cyan edges worn on the right side of her head (viewer's left, exactly as in <Picture 1>), brown leather boots and a leather satchel on a cross-body strap. <Picture 1> provides her outfit, proportions, boots, satchel and the position of her hair clip; <Picture 2> provides her exact face and eyes. summary: [reference generation] On the sunny hilltop meadow, <Subject 1> holds the small glass frame up against the sky, and inside the frame the golden afternoon becomes a starry night with a cyan-and-violet aurora and a distant floating glass city, keeping the exact appearance from <Picture 1> and the exact face from <Picture 2>. retention_analysis: <Subject 1> (appears in [Shot 1]): fully_preserved - face and violet-grey eyes from <Picture 2>; silver-white hair with cyan highlights, the glass hair clip on the right side of her head, the navy cloak with its luminous cyan-to-violet trim, boots and satchel from <Picture 1>. detailed_description: High-end anime film style matching cinematic key art: painterly backgrounds, soft rim light, warm low sunlight, crisp clean lineart and rich soft shading, not flat. [Shot 1] A medium close-up that starts just behind her right shoulder and slowly arcs around to a three-quarter front view of her face. <Subject 1>, hood down, holds the small flat glass frame up at arm's length above her eye line against the bright afternoon sky, so her face stays clear below it. Inside the rectangle of the frame the sky is completely different: a deep blue night full of stars with a slowly rippling ribbon of cyan and violet aurora, while everywhere outside the frame the sky stays golden with warm clouds. She moves the frame slowly from left to right and the night view inside it slides across the sky like a window, revealing the faint glowing spires of a floating glass city far away. Her violet-grey eyes widen in wonder and the cool aurora light shimmers softly on her face. Silver grass sways around her in the breeze. overall_soundscape: A warm breeze and rustling grass, then a low resonant hum and a soft shimmering chime when the night sky appears inside the frame, with a faint cold wind as if from far away. non_diegetic_music: N/A

    MiniMax H3dit5080-ref2va-fp8-pai8 (H3 Ref2VA · FP8 · PAI 8 steps)執行 9fc0fbe62026-10-19

  • AI 生成 · Powered by MiniMax H35.17 s · 124 f

    參考圖 → 影片

    穿過那道框 3・光之門

    ref2va-pc
    端到端
    169 秒
    解析度
    768×448
    種子
    6631
    • DGX Spark · GB1016.5 秒文字編碼
    • RTX 5080152 秒影片編碼 9.3DiT 120影片解碼 22.4
    • RTX 30602.7 秒音訊編碼 1.8音訊解碼 0.9
    • CPU0.3 秒封裝

    參考圖角色設定圖:The Courier+ 臉部特寫(左右鏡像)

    看參考圖
    臉部特寫(左右鏡像)AI 生成 · Qwen-Image 2.1
    臉部特寫(左右鏡像) · t2i‑pc · 執行 ff115974 · 種子 2026
    看完整 Prompt

    subject_definitions: <Subject 1> is the young woman shown in <Picture 1> (full-body reference) and <Picture 2> (head-and-shoulders close-up of the same woman), with long silver-white hair with faint cyan highlights, violet-grey eyes (soft lavender-grey irises), a dark navy hooded travel cloak edged with a thin luminous cyan-to-violet trim, a small hair clip of three overlapping flat translucent glass rectangles with thin glowing cyan edges worn on the right side of her head (viewer's left, exactly as in <Picture 1>), brown leather boots and a leather satchel on a cross-body strap. <Picture 1> provides her outfit, proportions, boots, satchel and the position of her hair clip; <Picture 2> provides her exact face and eyes. summary: [reference generation] On the golden meadow, the small glass frame that <Subject 1> releases swells into a tall glowing doorway onto a starry night, and she walks through it, keeping the exact appearance from <Picture 1> and the exact face from <Picture 2>. retention_analysis: <Subject 1> (appears in [Shot 1]): fully_preserved - face and violet-grey eyes from <Picture 2>; silver-white hair with cyan highlights, the glass hair clip on the right side of her head, the navy cloak with its luminous cyan-to-violet trim, boots and satchel from <Picture 1>. detailed_description: High-end anime film style matching cinematic key art: painterly backgrounds, soft rim light, warm low sunlight on one side and cool aurora light on the other, crisp clean lineart and rich soft shading, not flat. [Shot 1] A medium-wide shot from behind <Subject 1>, hood down, standing in the silver grass of the golden meadow, her whole figure in frame. She lets go of the small glass frame; it hangs in the air in front of her and, within the first second, swells smoothly into a tall door-sized rectangle of glowing glass edges standing in the grass. Through it lies a dark night shore under a vast cyan-and-violet aurora, while the meadow around it stays golden. A cool wind flows out of the doorway, pushing her silver-white hair and navy cloak back. She takes three calm steps forward and walks through the glowing doorway; as she crosses, the luminous trim of her cloak flares bright cyan. The camera follows behind her at walking pace and passes through the doorway with her into the night. overall_soundscape: A warm breeze over the grass, a deep airy whoosh as the frame swells into a doorway, a rising crystalline chord as she steps through, then sudden quiet with gentle waves on a distant shore. non_diegetic_music: N/A

    MiniMax H3dit5080-ref2va-fp8-pai8 (H3 Ref2VA · FP8 · PAI 8 steps)執行 2438ee522026-10-19

  • AI 生成 · Powered by MiniMax H35.17 s · 124 f

    參考圖 → 影片

    穿過那道框 4・碎光升空

    ref2va-pc
    端到端
    168 秒
    解析度
    768×448
    種子
    6641
    • DGX Spark · GB1016 秒文字編碼
    • RTX 5080151 秒影片編碼 10DiT 118影片解碼 22.4
    • RTX 30603.2 秒音訊編碼 2.3音訊解碼 0.9
    • CPU0.3 秒封裝

    參考圖角色設定圖:The Courier+ 臉部特寫(左右鏡像)

    看參考圖
    臉部特寫(左右鏡像)AI 生成 · Qwen-Image 2.1
    臉部特寫(左右鏡像) · t2i‑pc · 執行 ff115974 · 種子 2026
    看完整 Prompt

    subject_definitions: <Subject 1> is the young woman shown in <Picture 1> (full-body reference) and <Picture 2> (head-and-shoulders close-up of the same woman), with long silver-white hair with faint cyan highlights, violet-grey eyes (soft lavender-grey irises), a dark navy hooded travel cloak edged with a thin luminous cyan-to-violet trim, a small hair clip of three overlapping flat translucent glass rectangles with thin glowing cyan edges worn on the right side of her head (viewer's left, exactly as in <Picture 1>), brown leather boots and a leather satchel on a cross-body strap. <Picture 1> provides her outfit, proportions, boots, satchel and the position of her hair clip; <Picture 2> provides her exact face and eyes. summary: [reference generation] At night on a black rock shore under a vast aurora, <Subject 1> opens her hand and the small glass frame breaks softly into glowing shards that float up toward the aurora and a distant floating glass city, keeping the exact appearance from <Picture 1> and the exact face from <Picture 2>. retention_analysis: <Subject 1> (appears in [Shot 1]): fully_preserved - face and violet-grey eyes from <Picture 2>; silver-white hair with cyan highlights, the glass hair clip on the right side of her head, the navy cloak with its luminous cyan-to-violet trim, boots and satchel from <Picture 1>. detailed_description: High-end anime film style matching cinematic key art: night under a luminous aurora, cyan, blue and violet light reflected on a calm dark ocean, soft rim light, crisp clean lineart and rich soft shading, not flat. [Shot 1] A medium shot on the black rocks of a quiet shore at night, <Subject 1> framed from head to waist, hood down, the aurora rippling in bright bands across the starry sky behind her and the faint glowing spires of a floating glass city on the far horizon. She holds the small glass frame on her open palm at chest height and looks down at it with a calm smile; it breaks softly into dozens of tiny glowing glass shards that drift upward beside her, away from her face, turning slowly and flashing aurora light as they rise above her head. She lifts her gaze to follow them, the sea breeze moving her silver-white hair and cloak. The camera cranes up slowly from her shoulder, following the rising shards toward the aurora. overall_soundscape: Gentle waves lapping on the rocks, a soft sea breeze, and a delicate cascade of crystalline chimes as the shards rise and turn. non_diegetic_music: N/A

    MiniMax H3dit5080-ref2va-fp8-pai8 (H3 Ref2VA · FP8 · PAI 8 steps)執行 24ce157b2026-10-19

  • AI 生成 · Powered by MiniMax H35.17 s · 124 f

    參考圖 → 影片

    風鈴長廊

    ref2va-pc
    端到端
    167 秒
    解析度
    768×448
    種子
    6651
    • DGX Spark · GB1016 秒文字編碼
    • RTX 5080150 秒影片編碼 9.4DiT 118影片解碼 22.4
    • RTX 30602.5 秒音訊編碼 1.6音訊解碼 0.9
    • CPU0.3 秒封裝

    參考圖角色設定圖:The Courier+ 臉部特寫(左右鏡像)

    看參考圖
    臉部特寫(左右鏡像)AI 生成 · Qwen-Image 2.1
    臉部特寫(左右鏡像) · t2i‑pc · 執行 ff115974 · 種子 2026
    看完整 Prompt

    subject_definitions: <Subject 1> is the young woman shown in <Picture 1> (full-body reference) and <Picture 2> (head-and-shoulders close-up of the same woman), with long silver-white hair with faint cyan highlights, violet-grey eyes (soft lavender-grey irises), a dark navy hooded travel cloak edged with a thin luminous cyan-to-violet trim, a small hair clip of three overlapping flat translucent glass rectangles with thin glowing cyan edges worn on the right side of her head (viewer's left, exactly as in <Picture 1>), brown leather boots and a leather satchel on a cross-body strap. <Picture 1> provides her outfit, proportions, boots, satchel and the position of her hair clip; <Picture 2> provides her exact face and eyes. summary: [reference generation] <Subject 1> walks along a long wooden veranda hung with hundreds of clear glass wind chimes; a gust sweeps down the corridor and the chimes ring in a rolling wave toward her, keeping the exact appearance from <Picture 1> and the exact face from <Picture 2>. retention_analysis: <Subject 1> (appears in [Shot 1]): fully_preserved - face and violet-grey eyes from <Picture 2>; silver-white hair with cyan highlights, the glass hair clip on the right side of her head, the navy cloak with its luminous cyan-to-violet trim, boots and satchel from <Picture 1>. detailed_description: High-end anime film style, bright summer afternoon, dappled sunlight through green maple leaves, crisp clean lineart, rich painterly backgrounds and soft rim light. [Shot 1] A slow push-in along the long open wooden veranda of an old mountain temple, its eaves hung with hundreds of small clear glass wind chimes on short strings, each with a plain pale-blue paper tail, all hanging above head height. <Subject 1>, hood down, framed from head to knees, walks slowly toward the camera along the veranda. A gust of wind sweeps down the corridor from the far end: the chimes start swinging and ringing one after another in a rolling wave that travels toward her and past the camera, sunlight flashing through the glass and scattering small cyan and violet glints across the wooden floor and her cloak. She stops, closes her eyes and tilts her face up into the breeze with a small smile as her silver-white hair and cloak lift. The chimes never pass in front of her body or face. overall_soundscape: A rolling cascade of hundreds of glass wind chimes ringing close to the microphone, a soft gust of summer wind, cicadas in the trees, her light footsteps on the wooden boards. non_diegetic_music: N/A

    MiniMax H3dit5080-ref2va-fp8-pai8 (H3 Ref2VA · FP8 · PAI 8 steps)執行 f463432d2026-10-19

  • AI 生成 · Powered by MiniMax H35.17 s · 124 f

    參考圖 → 影片

    屋頂信使・躍上列車

    ref2va-pc
    端到端
    105 秒
    解析度
    768×448
    種子
    9061
    • DGX Spark · GB108.3 秒文字編碼
    • RTX 508096 秒影片編碼 5.7DiT 67.7影片解碼 22.6
    • RTX 30602.3 秒音訊編碼 1.4音訊解碼 0.9
    • CPU0.4 秒封裝

    用在上線預告・信使

    看完整 Prompt

    subject_definitions: <Subject 1> is the young woman shown in <Picture 1>, with long silver-white hair with faint cyan highlights, calm violet-grey eyes, a dark navy hooded travel cloak edged with a thin luminous cyan-to-violet trim, a small hair ornament of three stacked translucent glass frames, brown leather boots and a leather satchel on a cross-body strap. summary: [reference generation] <Subject 1> drops from a low rooftop onto the roof of a passing maglev train at night and steadies herself, keeping the exact appearance from <Picture 1>. retention_analysis: <Subject 1> (appears in [Shot 1]): fully_preserved - face, silver-white hair with cyan highlights, violet-grey eyes, the navy cloak with its luminous cyan-to-violet trim, the glass-frame hair ornament, boots and satchel. detailed_description: High-end 2D anime film style, night, light rain, neon city, crisp clean lineart, motion blur in the background. [Shot 1] A medium shot from the roof of a sleek white maglev train gliding fast through a neon city. <Subject 1> drops down from a low rooftop at the edge of the frame and lands on the flat train roof with bent knees, one hand down for balance. She rises carefully and finds her footing, her hood blown back and her silver-white hair and cloak snapping in the strong wind. Neon signs and lit windows streak past in long light trails behind her. The camera moves with the train, steady, at her height. overall_soundscape: A deep steady electric hum of the maglev, strong roaring wind, a metallic footstep on landing, light rain hitting the train roof. non_diegetic_music: N/A

    MiniMax H3dit5080-ref2va-fp8-pai8 (H3 Ref2VA · FP8 · PAI 8 steps)執行 a286077b2026-10-18

Prompt 裡的角色定義
<Subject 1> is the young woman shown in <Picture 1>, with long silver-white hair with faint cyan highlights, calm violet-grey eyes, a dark navy hooded travel cloak edged with a thin luminous cyan-to-violet trim, a small hair ornament of three stacked translucent glass frames, brown leather boots and a leather satchel on a cross-body strap.

從一張圖,到每天都在用的流程。

下面每一條流程,都只用到現在就有的功能,並連到操作導覽裡對應的章節。

  1. 分鏡 → 圖片 → 影片 → 時間軸 → MP4

    一個專案從頭做到成片。

    1. 在 Video Studio 的專案裡排好鏡頭,每個鏡頭一段 Prompt。
    2. 在 Image Studio 用 Qwen‑Image 2.1 生成每個鏡頭的圖,用 Takes 比較不同種子。
    3. 把圖當成鏡頭的 FL2VA 第一幀,生成有聲影片。
    4. 在時間軸上排列、切割、對齊,加上溶接、變速、調色與字幕。
    5. 用 FFmpeg 匯出一個 MP4(H.264 + AAC)。

    本頁的鏡頭是用控制器 API 跑同樣的管線(t2i‑pc → fl2va‑pc)生成的;短片是在 App 外用 FFmpeg 剪接,沒有用到 Video Studio 的時間軸。App 裡「圖庫圖片當第一幀」目前仍標示「尚未在 GPU 上驗證」。

    在操作導覽裡看
  2. 同一個角色,走過每一場戲

    一張參考圖,用在每個鏡頭。

    1. 在 Image Studio 生成角色設定圖,挑出一個種子。
    2. 用「Use in Video Shot…」把它設成鏡頭的 ref2va 參考圖。
    3. 在 Prompt 裡用 @Image1 指名她;Ref2VA 六段式範本會先依鏡頭的參考素材填好。
    4. 每個場景只改場景與動作,參考圖不變。
    5. Render 之後,在 Takes 裡並排比較不同種子。
    在操作導覽裡看
  3. 用 Prompt Advisor 反覆修 Prompt

    一邊打字,一邊對照官方格式。

    1. 一邊打字,74 條規則就在瀏覽器裡檢查 Prompt:45 條來自 MiniMax 官方指南,21 條來自已安裝的程式碼,8 條來自寫作指南。
    2. 其中 21 條附一鍵修正:先看差異預覽,按下 Apply 才會改。
    3. 錯誤等級的問題(例如 @ 名稱找不到對應素材)會在 Render 前擋下。
    4. 用同一個種子重新 Render,在 Takes 裡比較前後。

    Advisor 檢查的是格式與寫法;它對畫面品質的影響,還沒有用生成驗證過。

    在操作導覽裡看
  4. 讓 Agent 跑一整晚

    交給 Agent 排程,早上看比較。

    1. Agent 用 phasynctl(或可選的 MCP 伺服器,22 個工具)讀取狀態、啟動需要的元件。
    2. 每次 Render 前先 dry run,參數會對照即時的 schema 檢查。
    3. 控制器一次只跑一個 Run:Agent 逐一送出、等待、下載輸出。
    4. 早上在 Records 比較 2–8 筆紀錄(會列出干擾因素),再匯出 JSONL 或 CSV。
    在操作導覽裡看
  5. 把步驟分到自己的每台機器上

    哪一步在哪張卡跑,由你決定。

    1. 每台機器跑一個 NodeGateway,啟動後會自動向控制器註冊。
    2. 文字編碼、DiT、VAE、音訊、Mux 都是獨立行程,可以放在任何一台已加入、而且有對應執行環境的機器上。
    3. 在管線編輯器裡替每一步挑選元件實例,就能把它換到另一張卡。
    4. 大檔案(latent、影片)以 FileRef 在節點之間直接傳遞,不經過控制器。
    5. Records 會寫明每一步在哪台機器、花了多久。

    目前的執行環境只有 RTX 50 系列、RTX 30 系列與 GB10;新增機器要手動安裝。

    在操作導覽裡看

關於這些結果

  • 全部是 AI 生成:圖片來自 Qwen‑Image 2.1,影片與聲音來自 MiniMax H3,透過 Phasyn Studio 的控制器 API(t2i‑pc、fl2va‑pc、ref2va‑pc)在作者的機器上生成。「The Courier」是原創角色。
  • 模型授權:MiniMax H3 Community License(含地域限制與使用規範);Qwen‑Image 2.1 僅限研究與非商業用途。模型權重不隨專案提供。
  • 單次生成的結果,以及卡片上沒有標示放大倍率的短片,以原生解析度顯示,沒有放大:圖片 1152×672(角色設定圖 768×1024),影片 768×448、24 fps。網頁版的圖片轉成 WebP;單次生成的影片沿用原始 H.264 畫面串流,只把聲音做了響度正規化。
  • 時間是控制器觀察到的端到端時間與步驟時間,不含啟動或切換模型的時間(圖片 DiT 11.3 秒、dit‑5080 45.9 秒、Ref2VA DiT 40.6 秒)。
  • 不同 GPU 上的步驟可以同時進行(例如影片解碼與音訊解碼),所以步驟加總可能超過端到端時間。
  • 網頁版已做響度正規化;原始輸出約 −16~−51 LUFS、AAC 32 kHz。
  • 短片由多個生成鏡頭在 App 外用 FFmpeg 剪接而成(切點、溶接,部分鏡頭水平翻轉);標示「配樂:Suno」的短片,在 H3 自己的聲音之上墊了 Suno 生成的配樂(Music generated with Suno)。短片的混音已完成母帶處理(約 −14 LUFS),網頁版照原樣使用。
  • 卡片上標示放大倍率的作品(預告、幕後、《屋頂信使》與直式短片),把原生影片(768×448,直式生成為 448×768)放大到最多 1.75 倍,疊在同一鏡頭模糊、壓暗的背景上;只做重新取樣(lanczos 或 bicubic),沒有超解析,也沒有補幀。它們在 H3 自己的聲音下墊了 Suno 生成的配樂(Music generated with Suno),母帶約 −14 LUFS。
  • 部分畫面裡的招牌文字,是模型生成的偽文字,不是真實的店名或品牌。
即將開放原始碼與架構說明

想看它怎麼操作,還是想看數字?

操作導覽一步步走過真實的 App 畫面;效能實測寫明每一步跑在哪張卡(舊實驗封存沒記錄的標為「未記錄裝置」)。

看介紹影片