Open-Sora: 29K+ Stars — 开源视频生成完整安装指南 2026

# 核心推理流水线（简化版）
import torch
from opensora.models import STDiT3, T5Encoder, DC_AE
from opensora.schedulers import RectifiedFlowScheduler

# 加载模型
stdit = STDiT3.from_pretrained("hpcai-tech/Open-Sora-v2").cuda()
text_encoder = T5Encoder.from_pretrained("google/t5-v1_1-xxl").cuda()
vae = DC_AE.from_pretrained("hpcai-tech/Open-Sora-v2/vae").cuda()

# 编码提示词
prompt = "A serene underwater scene featuring a sea turtle"
prompt_embed = text_encoder.encode(prompt)  # [1, 512, 4096]

# 初始化潜变量噪声
latent = torch.randn(1, 16, 16, 128, 128).cuda()  # [B, C, T, H, W]

# 使用整流流去噪
scheduler = RectifiedFlowScheduler(num_steps=50)
for t in scheduler.timesteps:
    noise_pred = stdit(latent, t, prompt_embed)
    latent = scheduler.step(noise_pred, t, latent)

# 解码为视频
video = vae.decode(latent)  # [1, 3, 65, 768, 768]

# 创建虚拟环境
conda create -n opensora python=3.10 -y
conda activate opensora

# 克隆仓库
git clone https://github.com/hpcaitech/Open-Sora.git
cd Open-Sora

# 安装 PyTorch（CUDA 12.1）
pip install torch==2.4.0 torchvision==0.19.0 --index-url https://download.pytorch.org/whl/cu121

# 安装 Open-Sora
pip install -v .

# 安装可选加速器
pip install xformers==0.0.27.post2 --index-url https://download.pytorch.org/whl/cu121
pip install flash-attn --no-build-isolation

# 克隆仓库
git clone https://github.com/hpcaitech/Open-Sora.git
cd Open-Sora

# 构建 Docker 镜像
docker build -t opensora:latest .

# 运行交互式容器
docker run -ti --gpus all \
  -v $(pwd):/workspace/Open-Sora \
  -p 7860:7860 \
  --name opensora \
  opensora:latest

# 在容器内下载模型权重
pip install "huggingface_hub[cli]"
huggingface-cli download hpcai-tech/Open-Sora-v2 --local-dir ./ckpts

FROM nvidia/cuda:12.1.0-cudnn8-devel-ubuntu22.04

WORKDIR /workspace/Open-Sora

# 系统依赖
RUN apt-get update && apt-get install -y \
    git wget curl build-essential \
    libgl1 libglib2.0-0 \
    && rm -rf /var/lib/apt/lists/*

# Python 依赖
COPY requirements.txt .
RUN pip install -r requirements.txt

# 安装 Open-Sora
COPY . .
RUN pip install -v .

# 安装性能优化器
RUN pip install xformers==0.0.27.post2 --index-url https://download.pytorch.org/whl/cu121
RUN pip install flash-attn --no-build-isolation

EXPOSE 7860
CMD ["/bin/bash"]

# 方案 1：HuggingFace
pip install "huggingface_hub[cli]"
huggingface-cli download hpcai-tech/Open-Sora-v2 --local-dir ./ckpts

# 方案 2：ModelScope（适合中国用户）
pip install modelscope
modelscope download hpcai-tech/Open-Sora-v2 --local_dir ./ckpts

# 验证下载
ls -la ./ckpts/
# 预期输出：model.safetensors, config.json, vae/, text_encoder/

# 在独立环境中安装 ComfyUI
git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI
pip install -r requirements.txt

# 为 Open-Sora 创建自定义节点
mkdir -p custom_nodes/opensora-bridge
cd custom_nodes/opensora-bridge

# custom_nodes/opensora-bridge/opensora_node.py
import subprocess
import torch
import os

class OpenSoraTextToVideo:
    """ComfyUI 节点，用于 Open-Sora 文生视频生成"""
    
    @classmethod
    def INPUT_TYPES(cls):
        return {
            "required": {
                "prompt": ("STRING", {"multiline": True}),
                "resolution": (["256px", "768px"], {"default": "768px"}),
                "num_frames": ("INT", {"default": 65, "min": 17, "max": 129}),
                "steps": ("INT", {"default": 50, "min": 20, "max": 100}),
            }
        }
    
    RETURN_TYPES = ("VIDEO",)
    FUNCTION = "generate_video"
    CATEGORY = "video_generation"
    
    def generate_video(self, prompt, resolution, num_frames, steps):
        # 将提示词写入 CSV 用于批量处理
        with open("/tmp/opensora_input.csv", "w") as f:
            f.write(f"id,text\n0,\"{prompt}\"\n")
        
        # 启动推理
        cmd = [
            "torchrun", "--nproc_per_node", "1", "--standalone",
            "scripts/diffusion/inference.py",
            f"configs/diffusion/inference/{resolution}.py",
            "--save-dir", "/tmp/opensora_output",
            "--dataset.data-path", "/tmp/opensora_input.csv",
            "--num_frames", str(num_frames),
        ]
        subprocess.run(cmd, cwd="/workspace/Open-Sora")
        
        video_path = "/tmp/opensora_output/video_0.mp4"
        return (video_path,)

NODE_CLASS_MAPPINGS = {
    "OpenSoraTextToVideo": OpenSoraTextToVideo,
}

NODE_DISPLAY_NAME_MAPPINGS = {
    "OpenSoraTextToVideo": "Open-Sora Text to Video",
}

# configs/diffusion/inference/t2i2v_768px.py
# 文生图-图生视频配置
model = dict(
    type="STDiT3",
    from_pretrained="./ckpts/model.safetensors",
    enable_flash_attn=True,
    enable_layernorm_kernel=True,
)

vae = dict(
    type="DC-AE",
    from_pretrained="./ckpts/vae",
)

text_encoder = dict(
    type="t5",
    from_pretrained="./ckpts/text_encoder",
)

# FLUX T2I 模型配置
t2i_model = dict(
    type="flux-schnell",
    from_pretrained="black-forest-labs/FLUX.1-schnell",
)

# 采样配置
num_sampling_steps = 50
cfg_scale = 7.5
cfg_channel = 3  # 图像条件化尺度

# 安装 Gradio 依赖
pip install gradio spaces

# 启动 Web UI
python gradio/app.py --model-type v2 --checkpoint ./ckpts

# 安装 ColossalAI
pip install colossalai

# 多 GPU 训练启动（8x A100）
torchrun --nproc_per_node 8 --standalone \
    scripts/train.py \
    configs/diffusion/train/stage1.py \
    --data-path /path/to/video/dataset \
    --batch-size 2 \
    --mixed-precision bf16

# 长视频序列并行（>65 帧）
torchrun --nproc_per_node 8 --standalone \
    scripts/train.py \
    configs/diffusion/train/stage2_sp.py \
    --data-path /path/to/video/dataset \
    --sequence-parallel-size 4

# 1. CPU 卸载（节省 ~40% 显存，慢 25%）
torchrun --nproc_per_node 1 --standalone \
    scripts/diffusion/inference.py \
    configs/diffusion/inference/t2i2v_256px.py \
    --prompt "raining, sea" \
    --offload True

# 2. Flash Attention 3（加速 15-20%，无质量损失）
# 先安装：
git clone https://github.com/Dao-AILab/flash-attention
cd flash-attention/hopper
python setup.py install

# 3. 量化推理（INT8，节省 50% 显存）
torchrun --nproc_per_node 1 --standalone \
    scripts/diffusion/inference.py \
    configs/diffusion/inference/t2i2v_256px.py \
    --prompt "raining, sea" \
    --quantization int8

# 4. 混合精度（BF16）
torchrun --nproc_per_node 1 --standalone \
    scripts/diffusion/inference.py \
    configs/diffusion/inference/t2i2v_256px.py \
    --prompt "raining, sea" \
    --mixed-precision bf16

# 高分辨率生成的张量并行
torchrun --nproc_per_node 8 --standalone \
    scripts/diffusion/inference.py \
    configs/diffusion/inference/t2i2v_768px.py \
    --save-dir samples \
    --prompt "A soaring drone footage captures coastal cliffs" \
    --tp-size 4

# 有效提示词结构
prompt = """A cinematic wide shot of a golden retriever running along a sandy beach at sunset. 
Ocean waves break in the background with warm golden hour lighting. 
The camera follows the dog in slow motion, capturing flying fur details. 
High production value, anamorphic lens, shallow depth of field."""

# 避免：模糊或抽象的提示词
# 差："a dog video"
# 好：详细主体 + 动作 + 环境 + 光照 + 摄像机运动

# docker-compose.prod.yml
version: '3.8'

services:
  opensora:
    build: .
    runtime: nvidia
    environment:
      - NVIDIA_VISIBLE_DEVICES=all
      - CUDA_VISIBLE_DEVICES=0,1,2,3
      - HF_HOME=/workspace/cache
    volumes:
      - ./ckpts:/workspace/Open-Sora/ckpts:ro
      - ./samples:/workspace/Open-Sora/samples
      - huggingface_cache:/workspace/cache
    ports:
      - "7860:7860"
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    healthcheck:
      test: ["CMD", "python", "-c", "import torch; torch.cuda.is_available()"]
      interval: 30s
      timeout: 10s
      retries: 3
    restart: unless-stopped
    
  # 可选：批量任务队列工作器
  worker:
    build: .
    runtime: nvidia
    command: python scripts/diffusion/batch_worker.py --queue redis:6379
    environment:
      - NVIDIA_VISIBLE_DEVICES=4,5,6,7
    volumes:
      - ./ckpts:/workspace/Open-Sora/ckpts:ro
      - ./samples:/workspace/Open-Sora/samples
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 4
              capabilities: [gpu]

volumes:
  huggingface_cache:

# production_monitor.py
import torch
import time
import psutil
from prometheus_client import Counter, Histogram, start_http_server

# 指标
GENERATION_COUNTER = Counter('opensora_generations_total', '视频生成总数')
GENERATION_DURATION = Histogram('opensora_generation_seconds', '生成时间')
VRAM_USAGE = Histogram('opensora_vram_usage_bytes', '峰值显存使用')

def generate_with_monitoring(prompt, config):
    process = psutil.Process()
    start_mem = process.memory_info().rss
    
    torch.cuda.reset_peak_memory_stats()
    start_time = time.time()
    
    try:
        video = run_inference(prompt, config)
        
        duration = time.time() - start_time
        peak_vram = torch.cuda.max_memory_allocated()
        
        GENERATION_COUNTER.inc()
        GENERATION_DURATION.observe(duration)
        VRAM_USAGE.observe(peak_vram)
        
        return {
            'video': video,
            'duration': duration,
            'peak_vram_gb': peak_vram / 1e9,
            'peak_ram_gb': (process.memory_info().rss - start_mem) / 1e9,
        }
    except Exception as e:
        raise

# 在端口 9090 启动指标服务器
start_http_server(9090)