记录一下最近在本地折腾minimax h3的过程。
先直接上成片,点击在线观看:
5 个 MiniMax H3 片段
这次尝试的是5个仙侠or巨物?场景,每个大概7秒。最开始的提示词里写了推镜、跟拍、环绕,模型就真的在中途换镜头,画面看着挺热闹,但不是我想要的那种“像一张画自己活起来”。
后面把镜头改成完全固定,只让人物做一点点动作,主要让云、雾、瀑布、水面、光线和遗迹慢慢动。这样出来的画面更接近原图。
下面这5张是拿来做ref2va的原图和生视频提示词(原图gpt出的)

提示词 01|天门
Use <Picture 1> as the exact visual reference and preserve its opening composition. Preserve the same lone robed cultivator, costume, silhouette, reflective shallow water, monumental carved celestial gate, cliffside Chinese pavilions, waterfalls, storm clouds, color palette, lighting direction, horizon, and architectural geometry. This is one single continuous locked-off shot from a completely static tripod camera. The camera must remain fixed for the entire video. Keep the framing, crop, perspective, focal length, horizon, scale, and gate position unchanged. Do not pan, tilt, roll, zoom, dolly, push in, pull out, track, orbit, shake, reframe, recenter, change the camera angle, or follow the person. The person may make only subtle natural movements; the environment and background provide the main motion. No cuts, no shot changes, no scene changes, no new people, no creatures, no text, no modern objects. Keep the person full-body and recognizable throughout.
Timeline:
[0.0s-1.8s] Hold the exact opening composition. The person gently raises their head toward the glowing celestial gate and makes a small natural shift of weight. Only the hair, robe hem, shallow water, mist, clouds, and waterfalls move subtly.
[1.8s-5.7s] The person takes at most one slow, small step toward the gate and slightly raises one hand, then remains near the original position. Do not let the person run, turn around, or leave the frame. The fixed camera does not follow or recenter the person. The background changes gradually: cloud layers drift laterally through the gate, mist rolls across the lower cliffs, waterfalls continue flowing, warm light inside the arch slowly intensifies, and water ripples and reflections change naturally. Keep all buildings, cliffs, and the gate structurally unchanged.
[5.7s-7.3s] The person stops and holds the pose with only slight head, hand, hair, and robe movement. Clouds, mist, waterfalls, golden light, water ripples, and reflections continue changing. End on the same fixed composition and unchanged perspective.
Audio: no dialogue, no narration, no singing, no music. Only restrained natural ambience of wind, distant water, and very low environmental resonance; the final edit will replace this audio with one unified score. Negative motion constraints: camera movement, camera follow, subject tracking, automatic reframing, auto-centering, perspective shift, focal-length change, parallax from the camera, cuts, transitions, shot changes, scene replacement, architecture morphing, duplicated people, extra limbs, rubbery motion, flicker, warping, sudden zooms, and unstable faces.

提示词 02|浮空仙城
Use <Picture 1> as the exact visual reference and preserve its opening composition. Preserve the same lone robed figure on the rocky foreground at left, the immense floating Chinese immortal city, suspended mountain islands, bridges, waterfalls, lantern lights, cloud sea, giant celestial circular disk, original scale, color grading, mist, and architectural details. This is one single continuous locked-off shot from a completely static tripod camera. The camera must remain fixed for the entire video. Keep the framing, crop, perspective, focal length, horizon, scale, and city position unchanged. Do not pan, tilt, roll, zoom, dolly, push in, pull out, track, orbit, shake, reframe, recenter, change the camera angle, or follow the person. The person and the environment are the only moving elements. No cuts, no shot changes, no scene changes, no new people, no creatures, no text, no modern objects. Keep the person stable, recognizable, and grounded.
Timeline:
[0.0s-1.8s] Hold the exact opening composition. The person gently turns their head and upper body toward the floating city and makes a small natural shift of weight. Hair, robe, cloud sea, mist, and distant waterfalls move subtly.
[1.8s-5.7s] The person takes at most one slow, small step along the rocky foreground and slightly raises one sleeve toward the city, then remains near the original position. The fixed camera does not follow or recenter the person. The giant circular disk rotates slowly by only a few degrees, floating islands make a subtle heavy rise and fall, bridges sway almost imperceptibly, cloud layers drift laterally, waterfalls continue flowing, mist changes shape, and warm lanterns flicker gently. Keep every building and island structurally unchanged.
[5.7s-7.3s] The person holds the gesture with only slight sleeve and hair movement. The disk rotation eases, the city remains anchored, cloud sea and mist continue drifting, waterfalls continue flowing, and lanterns remain softly glowing. End on the same fixed composition and unchanged perspective.
Audio: clean, sparse xianxia background music. One airy, distant, ethereal adult female wordless vocalise with no lyrics, no spoken words, and no male voice, floating above restrained Chinese traditional instruments: xiao flute, guqin, guzheng, soft zhongruan, and very light temple bells. Slow spacious melody, delicate and mystical, clear stereo mix, gentle dynamics, no harsh transients. No drums, no beat, no modern synth, no EDM, no orchestra swell, no crowd, no dialogue, no sound effects, no wind noise, no water noise, no clipping, no distortion, and no excessive reverb. Negative motion constraints: camera movement, camera follow, subject tracking, automatic reframing, auto-centering, perspective shift, focal-length change, camera parallax, cuts, transitions, shot changes, scene replacement, architecture morphing, duplicated people, extra limbs, flicker, warping, and unstable faces.

提示词 03|玉剑遗迹
Use <Picture 1> as the exact visual reference and preserve its opening composition. Preserve the same lone small robed figure on the ancient stone bridge, the central monumental jade sword-like relic or sword-pillar, carved Chinese ruins, cliffs, waterfalls, dark water, storm light, and all original architectural geometry. Do not redesign the central relic into another object. This is one single continuous locked-off shot from a completely static tripod camera. The camera must remain fixed for the entire video. Keep the framing, crop, perspective, focal length, horizon, scale, bridge position, and relic position unchanged. Do not pan, tilt, roll, zoom, dolly, push in, pull out, track, orbit, shake, reframe, recenter, change the camera angle, or follow the person. The person and the environment are the only moving elements. No cuts, no shot changes, no scene changes, no new people, no creatures, no text, no modern objects. Keep the small figure anatomically stable and connected to the bridge.
Timeline:
[0.0s-1.8s] Hold the exact opening composition. The person gently raises their head toward the jade relic and makes a small natural shift of weight. Robe, hair, mist, waterfalls, and dark water move subtly.
[1.8s-5.7s] The person takes at most one slow, small step toward the relic and slightly raises one hand, then remains near the original position. The fixed camera does not follow or recenter the person. The jade relic performs one restrained, small vertical rise-and-settle motion without changing shape. Faint carved jade patterns illuminate gradually from the base upward and then remain softly glowing. Mist drifts across the bridge, waterfalls flow, storm clouds shift, and the dark water reflection changes naturally. Keep the ruins, cliffs, bridge, and relic geometry unchanged.
[5.7s-7.3s] The person stops and holds their gaze. The relic settles completely while its faint patterns continue glowing; mist, waterfalls, storm light, and water reflections continue changing subtly. End on the same fixed composition and unchanged perspective.
Audio: clean, sparse xianxia background music. One airy, distant, ethereal adult female wordless vocalise with no lyrics, no spoken words, and no male voice, floating above restrained Chinese traditional instruments: xiao flute, guqin, guzheng, soft zhongruan, and very light temple bells. Slow spacious melody, delicate and mystical, clear stereo mix, gentle dynamics, no harsh transients. No drums, no beat, no modern synth, no EDM, no orchestra swell, no crowd, no dialogue, no sound effects, no wind noise, no water noise, no clipping, no distortion, and no excessive reverb. Negative motion constraints: camera movement, camera follow, subject tracking, automatic reframing, auto-centering, perspective shift, focal-length change, camera parallax, cuts, transitions, shot changes, scene replacement, relic redesign, architecture morphing, duplicated people, extra limbs, flicker, warping, and unstable faces.

提示词 04|天钟圣境
Use <Picture 1> as the exact visual reference and preserve its opening composition. Preserve the same lone robed figure on the left cliff, the enormous suspended bronze celestial bell, ancient Chinese sanctuary, waterfalls, cliffs, mist, clouds, warm light beneath the bell, and the original architecture and proportions. Keep the bell massive, solid, and physically grounded in the air. This is one single continuous locked-off shot from a completely static tripod camera. The camera must remain fixed for the entire video. Keep the framing, crop, perspective, focal length, horizon, scale, bell position, and figure position unchanged. Do not pan, tilt, roll, zoom, dolly, push in, pull out, track, orbit, shake, reframe, recenter, change the camera angle, or follow the person. The person and the environment are the only moving elements. No cuts, no shot changes, no scene changes, no new people, no creatures, no text, no modern objects.
Timeline:
[0.0s-1.8s] Hold the exact opening composition. The person makes at most one small step toward the cliff edge and gently raises their head toward the bell. Robe, hair, mist, clouds, and waterfalls move subtly.
[1.8s-5.7s] The person stops and remains near the original position, making only a slight natural hand or shoulder movement. The fixed camera does not follow or recenter the person. The bell makes one very small, slow forward-and-back swing and a restrained visual resonance: the bronze surface vibrates subtly, the warm light beneath the bell pulses once, and a soft circular disturbance moves through the mist. Waterfalls continue flowing, cloud layers drift, and the sanctuary remains structurally unchanged.
[5.7s-7.3s] The person holds the upward gaze. The bell eases to stillness while its warm glow remains; mist, clouds, waterfalls, and light continue changing gently. End on the same fixed composition and unchanged perspective.
Audio: clean, sparse xianxia background music. One airy, distant, ethereal adult female wordless vocalise with no lyrics, no spoken words, and no male voice, floating above restrained Chinese traditional instruments: xiao flute, guqin, guzheng, soft zhongruan, and very light temple bells. Slow spacious melody, delicate and mystical, clear stereo mix, gentle dynamics, no harsh transients. No drums, no beat, no modern synth, no EDM, no orchestra swell, no crowd, no dialogue, no sound effects, no wind noise, no water noise, no clipping, no distortion, and no excessive reverb. Negative motion constraints: camera movement, camera follow, subject tracking, automatic reframing, auto-centering, perspective shift, focal-length change, camera parallax, cuts, transitions, shot changes, scene replacement, bell redesign, architecture morphing, duplicated people, extra limbs, flicker, warping, and unstable faces.

提示词 05|环形遗迹宫殿
Use <Picture 1> as the exact visual reference and preserve its opening composition. Preserve the same lone robed figure in the foreground by the water, the colossal engraved circular ring relic, Chinese mountain palaces, cliffs, waterfalls, lake surface, sunset clouds, reflections, color palette, and original geometry. Keep the person full-body and the ring colossal and stable. This is one single continuous locked-off shot from a completely static tripod camera. The camera must remain fixed for the entire video. Keep the framing, crop, perspective, focal length, horizon, scale, ring position, figure position, and reflection alignment unchanged. Do not pan, tilt, roll, zoom, dolly, push in, pull out, track, orbit, shake, reframe, recenter, change the camera angle, or follow the person. The person and the environment are the only moving elements. No cuts, no shot changes, no scene changes, no new people, no creatures, no text, no modern objects.
Timeline:
[0.0s-1.8s] Hold the exact opening composition. The person gently raises their head toward the giant ring and makes a small natural shift of weight. Robe, hair, lake ripples, mist, waterfalls, and reflections move subtly.
[1.8s-5.7s] The person takes at most one slow, small step toward the shoreline and slightly raises one hand toward the ring, then remains near the original position. The fixed camera does not follow or recenter the person. The colossal ring rotates slowly around its vertical axis by only a few degrees; engraved patterns illuminate in a restrained sequence. Clouds drift through the ring opening, waterfalls flow, mist changes shape, and the lake reflection develops soft expanding ripples. Keep the ring, palaces, cliffs, and architecture structurally unchanged.
[5.7s-7.3s] The person stops and holds the gesture with only slight hand, hair, and robe movement. The ring rotation eases to stillness, its warm rim light remains, clouds and mist continue drifting, waterfalls continue flowing, and the reflection settles gradually. End on the same fixed composition and unchanged perspective.
Audio: clean, sparse xianxia background music. One airy, distant, ethereal adult female wordless vocalise with no lyrics, no spoken words, and no male voice, floating above restrained Chinese traditional instruments: xiao flute, guqin, guzheng, soft zhongruan, and very light temple bells. Slow spacious melody, delicate and mystical, clear stereo mix, gentle dynamics, no harsh transients. No drums, no beat, no modern synth, no EDM, no orchestra swell, no crowd, no dialogue, no sound effects, no wind noise, no water noise, no clipping, no distortion, and no excessive reverb. Negative motion constraints: camera movement, camera follow, subject tracking, automatic reframing, auto-centering, perspective shift, focal-length change, camera parallax, cuts, transitions, shot changes, scene replacement, ring redesign, architecture morphing, duplicated people, extra limbs, flicker, warping, and unstable faces.
最后放一下本机配置。
系统:Windows
GPU:NVIDIA GeForce RTX 5060 Ti 16GB
内存:约 48GB
NVIDIA 驱动:610.88(较大可能解决gpu lost问题)
ComfyUI:0.31.0
PyTorch:2.10.0+cu130
comfy-kitchen:0.2.28
comfy-aimdo:0.4.13
本次工作流(直接就官方给的):
MiniMaxH3ReferenceToVideo
Ref2VA Turbo
Turbo LoRA strength:1.0
low_vram:false
scheduler:simple
steps:6
内部尺寸:1280x736
最终尺寸:1280x720
帧率:24 fps
帧数:175
实际时长:约 7.292 秒
视频:H.264
音频:AAC,32 kHz,双声道
这5条最后用的都是ref2va,不是fl(貌似ref效果更好一点?)。原图2560×1440,ref的match会自己适配参考图,本地原本生成1280×736,最后裁成1280×720,没有再接一轮upscale(效用不高)。单条大概12分钟,5条连续跑下来差不多一个小时。期待后续社区进一步优化。
最后一点碎碎念,5060ti买来大半年了,平时不怎么玩游戏一直吃灰,现在算是开发了一个新玩法,感觉minimaxh3就是本地生视频deepseek时刻,期待后续国产开源模型越来越好,本地做大电影不是梦。