Perflow-Shuai/MiniMax-H3-3Step-LoRA

MiniMax-H3 3-Step

Ready-to-use merged H3 transformer weights. No LoRA loading or merging is required.

中文：LoRA 已合入主模型。直接加载本仓库的 transformer/，无需再挂 LoRA。

Generates 1344×768 video with 32 kHz stereo audio: 124 frames at 24 FPS. Uses 3 evaluations per chunk, 8 chunks (24 evaluations per clip), CFG 1. Use the included chunked sampler, not a stock H3 pipeline with only its step count changed.

Quick start

Use Python 3.12 and a compatible CUDA-enabled PyTorch environment. The tested environment uses PyTorch 2.11.0/CUDA 13.0; requirements.txt pins the Diffusers revision. A high-memory GPU is required; consumer-GPU offloading and quantization are not validated.

hf download Perflow-Shuai/MiniMax-H3-3Step-LoRA --local-dir ./h3-3step
hf download MiniMaxAI/MiniMax-H3 \
  --revision fa9c8ab1eaa21c8ae25e7e40b83b2e6002f340af \
  --local-dir ./h3-base --exclude "transformer/*"
cd h3-3step
pip install -r requirements.txt
python infer.py --base ../h3-base --prompt-file examples/tokyo-car.txt \
  --output ./output --seed 20262020

Output: output/video.mp4, including generated audio. The script automatically loads the complete merged transformer from this repository. Its approximately 66.28 GB of weights are stored as 14 Safetensors shards plus an index; these shards form one full transformer checkpoint, not separate LoRAs. No standalone adapter weights are included or loaded. Use a fresh local directory if upgrading. The original H3 download supplies the unchanged text encoder, VAEs and supporting components. Replace the example prompt with your own UTF-8 text file. Rerun the same command/output directory to resume an interrupted clip.

中文：下载本仓库和 H3 的非 transformer 组件后，直接执行上述推理命令即可。视频下方附完整英文 Prompt 和中文翻译；实际输入使用英文。

Python integration

from h3_inference import load_base, ChunkedH3Runtime, ChunkGeometry, generate_sample

model = load_base("/path/to/h3-3step")
runtime = ChunkedH3Runtime(model, checkpoint_blocks=False)
video_rows, audio_rows = generate_sample(
    runtime, ChunkGeometry(), {"prompt_embeds": prompt_embeds},
    seed=20262020, nfe=3,
)

prompt_embeds is the positive H3 text-encoder output before the token refiner, shape [1, text_length, text_dim]. See infer.py for text encoding and VAE decoding, and inference_config.json for the chunk boundaries and paired audio/video schedules. The current interface is first-clip text-to-audio/video, not long-video continuation.

Video cases

These videos were generated with the merged transformer and the included 3-step-per-chunk sampler. Native BF16/FP32 weights are retained. Merging introduces floating-point rounding, so outputs can change in composition and are not pixel-identical to separate FP32 adapter inference.

Tokyo night drive / 东京夜间行车

Seed: 20262020 · Video · Prompt file

English prompt

integrated_multimodal_description: [Shot 1] A photoreal street-racing night scene in Tokyo opens low and wide on a turbocharged wide-body hot hatch blasting through a wet expressway lane, sodium lights streaking across its flared fenders. The camera tracks inches above the asphalt beside the car as the turbo spools, the driver snaps a shift, and blue reflections from overhead signs slide across the windshield; halfway through, the car dives under an overpass and the camera arcs slightly behind to reveal the aggressive rear wing and a brief flame popping from the exhaust before it disappears into the neon distance.

overall_soundscape: Tire hiss on damp pavement, rising turbo whine, sharp gear-change clacks, and a deep exhaust note bounce off concrete walls. Passing traffic rushes by in the far lanes and the city hum stays constant underneath.

non_diegetic_music: N/A

中文翻译

综合多模态描述（integrated_multimodal_description）：[镜头 1] 一段照片般真实的东京街头赛车夜景，以低机位广角画面开场：一辆加装涡轮增压和宽体套件的高性能掀背车，在湿漉漉的高速公路车道上飞驰，钠灯光线掠过它外扩的翼子板。摄影机贴着车身侧面，在距离沥青路面仅数英寸的高度跟拍；涡轮转速攀升，驾驶员迅速换挡，头顶路牌的蓝色倒影滑过挡风玻璃。镜头进行到一半时，汽车冲入立交桥下方，摄影机略微向后绕行，展现夸张的尾翼以及排气管中短暂喷出的火焰，随后汽车消失在远处的霓虹灯光中。

整体声景（overall_soundscape）：轮胎擦过潮湿路面发出嘶嘶声，涡轮的尖啸逐渐升高，清脆的换挡咔嗒声与低沉的排气声在混凝土墙面之间回荡。远处车道上有车辆呼啸而过，城市的低鸣始终持续在这些声音之下。

非叙事音乐（non_diegetic_music）：不适用（N/A）。

Cartoon recording studio / 卡通录音室

Seed: 20262022 · Video · Prompt file

English prompt

integrated_multimodal_description: [Shot 1] A high-end cartoon recording studio scene opens in a polished family-animation style on a songwriter seated at a desk beneath acoustic panels, headphones around the neck, a notebook open beside a glowing mixing console. In a medium close-up, the camera drifts from the console to the character's hand as they tap the pencil once in thought, then begin writing fresh lyrics in smooth realistic strokes across the page while their other hand lightly keeps time on the tabletop; by the end, they pause, smile at the line they have finished, and underline it.

overall_soundscape: A faint studio air hum, the soft buzz of powered equipment, pencil scratches on paper, and light fingertip taps create an intimate working rhythm. The swivel chair creaks once as the writer leans in to finish the line.

non_diegetic_music: N/A

中文翻译

综合多模态描述（integrated_multimodal_description）：[镜头 1] 一个高品质的卡通录音室场景，以精致的合家欢动画风格开场：一位词曲作者坐在吸音板下方的桌前，耳机挂在脖子上，一本打开的笔记本放在发光的混音台旁边。在中近景中，摄影机从混音台缓缓移向角色的手：作者思考时用铅笔轻敲一下，随后开始用流畅而逼真的笔画在纸上写下新的歌词，另一只手则在桌面上轻轻打拍子。镜头结束时，作者停下笔，对着刚写好的那一行微笑，并在其下方画线。

整体声景（overall_soundscape）：录音室里微弱的空气低鸣、通电设备轻柔的嗡嗡声、铅笔划过纸面的沙沙声，以及指尖轻敲桌面的声音，共同形成亲密的工作节奏。作者俯身写完那一行时，转椅吱呀响了一次。

非叙事音乐（non_diegetic_music）：不适用（N/A）。

Morning river / 清晨河景

Seed: 21262023 · Video · Prompt file

English prompt

integrated_multimodal_description: [Shot 1] A clean looping landscape shot looks across the Moscow River from a fixed embankment viewpoint under pale morning light, with city buildings softened in the distance. The water surface continuously ripples past in overlapping silver-blue patterns, reflecting slow-moving clouds and occasional glints from the skyline while the frame remains calm and seamless for a perfect loop.

overall_soundscape: Gentle water movement laps against the embankment, distant traffic murmurs across the river, and a faint breeze moves through the open air. No abrupt sounds interrupt the steady urban calm.

non_diegetic_music: N/A

中文翻译

综合多模态描述（integrated_multimodal_description）：[镜头 1] 一个干净、可循环播放的风景镜头，从河堤上的固定视点望向莫斯科河，清淡的晨光之下，远处城市建筑的轮廓显得柔和。水面持续泛起交叠的银蓝色涟漪，映出缓缓移动的云朵和城市天际线偶尔闪现的亮光；画面始终平静，首尾无缝衔接，构成完美的循环。

整体声景（overall_soundscape）：轻柔的水流拍打河堤，河对岸传来远处交通的低鸣，微风掠过开阔的空间。没有突兀的声音打破这份持续的城市宁静。

非叙事音乐（non_diegetic_music）：不适用（N/A）。

Crystal heart / 水晶心

Seed: 20262027 · Video · Prompt file

English prompt

integrated_multimodal_description: [Shot 1] A luminous fantasy landscape opens on a bright sunny sky where a crystal-clear cloud shaped like a heart hangs motionless above a dense green forest. The camera drifts upward from the treetops to frame the heart cloud against deep blue, and sunlight refracts through it like cut glass, scattering tiny rainbow flecks across the canopy before the shape slowly sways in the high breeze.

overall_soundscape: A gentle forest ambience of birdsong and distant leaves in the wind fills the lower world, while the open sky above feels airy and quiet. The mood stays calm and radiant.

non_diegetic_music: N/A

中文翻译

综合多模态描述（integrated_multimodal_description）：[镜头 1] 一个明亮的奇幻风景，以阳光灿烂的天空开场：一朵晶莹剔透、形状如心的云，静静地悬在浓密的绿色森林上方。摄影机从树梢缓缓向上移动，将心形云衬在深蓝色天空前；阳光像穿过切割玻璃一样，在云中折射，将细小的彩虹光斑洒向林冠。随后，心形轮廓在高空微风中缓缓摇曳。

整体声景（overall_soundscape）：下方的森林中，鸟鸣与远处树叶随风作响，构成轻柔的环境声；上方开阔的天空则显得轻盈而安静。整体氛围始终平和、明亮。

非叙事音乐（non_diegetic_music）：不适用（N/A）。

Terms

Built with MiniMax-H3. These are modified transformer weights, not the unchanged official base checkpoint. The MiniMax-H3 Community License Agreement, including its territory restrictions and acceptable-use requirements, applies. See NOTICE. Inference code is covered by CODE_LICENSE. Text encoder and VAE weights are not redistributed here.