ローカルのMiniMax H3で、約9分で15秒の2K超え動画を生成する話(RTX 5090)。 Alibabaの蒸留LoRAで480pの当たりを引いて、同じlatentを1344×768に拡大して4ステップのrefine、最後にRTX Video Super Resolutionで2倍の2688×1536にアップスケール、という分担です。 RTX VSRは、拡散モデルのように描き直すのではなくフレームを1回通すだけなので、2688×1536へのアップスケールでも1フレーム0.1〜0.2秒くらいで済みます。 ただし、再生成ではないので元の絵がダメならダメなまま大きくなります。今回もrefineで絵を綺麗にしてから解像度を上げていますが、やはり2Kといっても寝ぼけた感じにはなってしまっています。 数値は今日の実測(362フレームで約85秒、モデル読み込み含む)を丸めたものです。 480p生成が約2分55秒、refineが約4分40秒、VSRが約1分25秒。 画質だけなら非蒸留の25ステップや、Alibabaの蒸留LoRAで最初から768pで回したほうがいいのかもしれないです。 ただ、480pなら1本3分弱で回して、当たりだけ、保存しておいたlatentでアップスケールすればいいので、こちらのほうが気楽に生成できます。

参照に使った画像と、最後に使ったプロンプトです。 部屋の絵1枚、キャラの体3面と頭3面、配信アバター1枚の計8枚。部屋の絵は素のまま渡して、モニターに映るアバターの配置・動きは参照だけでH3に任せています。 手順のメモはGitHubに置いています。 （以下、プロンプト全文） subject_definitions: <Subject 1> is the streaming room and its camera viewpoint, taken from <Picture 1>: a high-angle view looking down from near the ceiling into a small messy apartment room at night, a large white L-shaped desk in the right half with three monitors, a webcam on top of the center monitor, a microphone on a boom arm, a keyboard and a mouse on a pink desk mat, plastic bottles and snack packages piled on the desk and rolling on the wooden floor, a grey gaming chair in front of the desk, an unmade single bed with a cream comforter in the lower left, pink curtains half drawn over a window showing a dark night sky, a closet on the far left, the room brightly lit by a warm ceiling light. <Subject 2> is the young woman, defined by <Picture 2>, <Picture 3> and <Picture 4> (body: front, left side, back) and <Picture 5>, <Picture 6> and <Picture 7> (head: front, left profile, back): long black hair tied up in a loose messy bun with strands falling out and a lavender ribbon tied in the bun, large violet eyes with heavy lids and long lashes, pale porcelain skin, no makeup, a cool detached expression, slim build. She wears an oversized matte black zip-up tracksuit jacket with a hood, lavender stripes down the sleeves, thin white frilled lace along the inner collar edge on both sides, a lavender ribbon bow on the chest, a small ribbon on the zipper pull, the jacket half zipped over a plain white t-shirt and sagging off one shoulder, loose black track pants with lavender stripes, elastic ankles bunched up, and bare feet. <Subject 3> is her 2D streaming avatar and its stream background, defined by <Picture 8>: a cute bust-up anime girl with fluffy wavy hair fading from white to glowing mint like a jellyfish, large sparkling eyes, big over-ear headphones with jellyfish-shaped ornaments, a dark capelet over a white top, smiling brightly and waving, standing in a glowing virtual laboratory with tall glass tubes, neon-lit shelves and floating bubbles in cyan and pink behind her; this whole picture, avatar plus lab background, is what the center-right monitor shows as a flat 2D stream image. summary: [reference generation] A single locked high-angle shot of <Subject 2> ending her stream in <Subject 1>: she waves goodbye to the webcam with her back to us in a bright, high-energy streamer voice while her 2D avatar on the monitor mirrors the wave, clicks the mouse to stop the stream, slumps in silence, picks up a short black heated tobacco device with a white stick from the desk, spins the chair around fast so her face is seen for the first time, slides down in the chair and throws her bare legs onto the bed, draws one deep, slow breath through it with the air around her face staying completely clear, holds the breath, and finally lets it out in a long exhale at the very end. retention_analysis: <Picture 1>: fully_preserved - the room layout, furniture, clutter, lighting and the exact camera position and angle are kept; the woman is added in the chair, and while the stream is on the center-right monitor shows <Subject 3> instead of the plain dashboard seen in the picture. <Picture 2>: fully_preserved - body proportions and outfit. <Picture 3>: fully_preserved - outfit from the side. <Picture 4>: fully_preserved - back of the jacket, hair bun and ribbon from behind. <Picture 5>: fully_preserved - face and hairstyle. <Picture 6>: fully_preserved - profile of the face and bun. <Picture 7>: fully_preserved - back of the head and bun. <Picture 8>: fully_preserved - the avatar and its background shown on the monitor keep this exact design, colors and proportions in every frame while the stream is on. detailed_description: The entire target video is rendered in the exact clean cel-shaded anime style of <Picture 1> and <Picture 2>: crisp lineart, flat pastel fills with soft two-tone shadows, limited-animation clarity. <Picture 1> provides the set and the camera: the camera is locked in the same high-angle position for the whole video, Static Shot, no panning, no tilting, no zoom, no cut. [Shot 1] A single continuous 15-second take. At the start the woman sits in the grey gaming chair facing the desk with her back to the camera, only the back of her black jacket and the messy bun with the lavender ribbon visible, the monitors in front of her showing a streaming dashboard and chat that stay exactly the same, pixel-still, for the entire video, and on the center-right monitor the streaming avatar <Subject 3> exactly as in <Picture 8>, crisp and unchanging in design. A heated tobacco device lies on the desk just to the left of the mouse: a short matte black cylinder about the size of her thumb, with a tiny rectangular display and two small buttons on its face, and a short white tobacco stick standing out of its top. At 00:00.000 she leans toward the webcam on the center monitor and throws both arms up high, waving both hands big and fast above her head, saying her sign-off in a bright, bubbly, high-pitched streamer voice, full of energy and rising at the end of each phrase: (S1) <d>[Japanese] 今日はここまで〜！ばいば〜い！</d> On the center-right monitor, the avatar <Subject 3> (the same jellyfish-haired girl from <Picture 8>, same face, same colors, same headphones, same neon laboratory behind her) raises both arms and waves both hands at exactly the same moments, mirroring every swing of her real arms and every bob of her head frame for frame, like a live-tracked 2D model; its design never changes. Her face stays turned away from us toward the webcam. She keeps both hands waving for a full second after the last word of the line is spoken, still facing the webcam, the avatar waving with her. At 00:03.500, well after she has finished the whole line and stopped waving, she brings her right arm down in a clear visible arc to the desk, reaches across the pink desk mat and places her right hand on the white mouse, her fingers wrapping over it in plain view of the camera; she nudges the mouse a few centimeters to move the cursor, then presses the left button with her index finger with a distinct small click motion, and keeps her hand resting on the mouse for a moment before letting go. Only the center-right monitor reacts; the left monitor and the right monitor keep showing exactly the same dashboard and chat image as at the start, with no flash, no white screen, no scrolling and no layout change. At 00:04.500 the moment the stream is off, her whole posture changes: the bright streamer smile switches off, her shoulders drop, her head hangs forward for a moment and she lets out a long, heavy sigh with no words, the exhaustion shown only in her posture. The avatar <Subject 3> disappears from the monitor, replaced by a plain idle screen. From this moment to the end of the video, the center-right monitor shows only a plain blank idle screen, and the left and right monitors continue to show the identical static dashboard and chat image they showed from the first frame, frozen and unchanged; no face, no character, no avatar of any kind appears on any screen again. At 00:05.500 she looks down at the desk, reaches her right hand across to the short black heated tobacco device lying on the desk mat just to the left of the mouse, wraps her fingers around the black cylinder with the white stick pointing up, and lifts it clearly off the desk, then holds it loosely in her right hand; this pick-up is shown plainly, the black device with its white stick visible in her hand as she brings it back. At 00:07.000 the chair spin is timed like Japanese limited animation, with a clear anticipation, a snap and a held pose: first she plants one bare foot against the desk and coils her upper body a little in the opposite direction, holding that wound-up pose completely still for a beat of about half a second (the tame); then she kicks off hard and the chair whips around toward the camera in only a few frames, a fast blurred half turn with her hair and the lavender ribbon streaking behind her (the tsume); at the same moment as the spin she raises her left hand and sweeps it back through her bangs, pushing her hair up and off her forehead in one lazy stroke as her face comes around, the hair falling back loosely afterward; the chair overshoots slightly, snaps back, and freezes in a held pose facing the camera, her left hand still resting in her hair for a beat, while only her hair strands and the loose jacket keep swinging and settle late. As the chair whips around her face is seen for the first time: heavy-lidded violet eyes, a calm detached expression. At 00:08.500 she slides down in the chair until she is half lying in it, her back low against the backrest, and throws both bare legs out to her left onto the bed, heels landing on the edge of the mattress, the track pants riding up to show her ankles. At 00:09.500 she raises the device so the white stick touches her lips and takes a very long, very slow inhale from it that lasts a full two and a half seconds, until 00:12.000, the black cylinder held in her fist below her mouth, her chest rising slowly the whole time, eyes gradually closing; it is a deep inhale, drawn in through the device: her chest and shoulders rise, her lips stay sealed around the tip, and the air around her face, the device and the whole room stays perfectly clear and clean the entire time, exactly like the air in every earlier second of the video; the only sounds during the inhale are a faint, slow intake of breath and the room tone. At 00:12.000 she lowers the device and holds the breath with her lips closed and her eyes shut, completely still, not breathing out at all, for a long two seconds until 00:14.000; during this hold her lips stay sealed, her chest stays full, and the air around her face remains perfectly clear and clean, identical to the air during the inhale; this hold is silent apart from the PC fan hum and the room tone, with her breath completely held. At 00:14.000 she tilts her head back and only now, for the first time, begins to exhale, a slow, long, drawn-out breath that continues all the way to the end of the video, its soft exhale sound starting at exactly this moment and lasting as long as the visible breath, and only with this exhale, after 00:14.000, does anything visible leave her lips: a faint, thin, barely visible wisp of pale vapor that dissolves within a second; her head resting back against the chair as the video ends on the same locked wide shot. Movement is unhurried throughout with easing on every start and stop, and every action uses Japanese anime timing: a short anticipation, a fast decisive main motion, then a held pose with only hair and clothing settling; hair strands and the loose jacket settle after each motion. The bottles, clothes and clutter in the room stay exactly where they are. The air in the room is perfectly clear and clean for the whole video, including the pick-up, the deep inhale and the breath-hold; the single exception is the exhale after 00:14.000, a faint thin wisp only, gone within a second. Consistent character identity, hairstyle, outfit and cel-shaded anime style throughout; no morphing, no identity drift, no extra people, no second character, no other avatar or face on any screen after the stream ends, no camera movement, no cuts, no glow, no retro anime look, no photorealistic rendering. overall_soundscape: her bright high-pitched energetic streamer voice for the sign-off line, then a single mouse click, a long heavy wordless sigh, the squeak and rattle of the office chair spinning fast, a soft rustle as she pushes her hair back, the faint hum of PC fans, silence with only the fan hum while she inhales slowly and holds her breath, then a soft, long, slow exhale sound that begins only at 00:14.000 in sync with the visible breath and lasts to the end, quiet apartment room tone at night. non_diegetic_music: N/A

minimax h3的开源协议限定限制了美国 欧盟 韩国的本地部署，很神奇的对日本没有任何限制呢

Thanks for sharing! My workflow is pretty much the same. I also use 0.4mp for the draw and when I see it looks decent, I zoom in on the space. I just went straight to 1.8mp instead _(:з」∠)_

slopu

RTX 5090でH3ローカル生成9分で2K超え、蒸留LoRA→refine→RTX VSRの分担が効率的ですね。H3プロンプト例をオープンソースで公開中：

DLSS 5 with frame gen my jigga

好厉害

这个流程很有意思，等于把生成分成三份工作：先用蒸馏LoRA快速拿到低分辨率基础，再用latent放大和refine控制细节，最后靠RTX的视频超分补足最终分辨率。这样既避免了全程跑高分辨率的高显存压力，也把5090的硬件优势用在了最后一步。

ローカルの Minmax h3。用途を限定すればめちゃくちゃ使いやすいんですよね。サーバーの Minmax でもやっぱり 1 分 100 円はかかるので。24時間で 10 万円越え。。。うーん GeForce 5080 買おうかなあ。。。