全球首款无限时长电影生成模型SkyReels-V2本地部署教程：视频时长无限制！

一、简介

SkyReels-V2 模型集成了多模态大语言模型（MLLM）、多阶段预训练、强化学习以及创新的扩散强迫（Diffusion-forcing）框架，实现了在提示词遵循、视觉质量、运动动态以及视频时长等方面的全面突破。通过扩散强迫框架和 多阶段优化技术 ，首次实现了单镜头 30 秒、40 秒的流畅输出，并通过“ Extend ”无限延伸，彻底打破了时长枷锁。

二、环境部署

基础环境要求:

环境名称	版本信息 1
Ubuntu	22.04.5 LTS
Cuda	V12.4.131
Python	3.12.7
NVIDIA Corporation	RTX 4090

从算家云基础镜像开始创建：

三、SkyReels-V2安装

首先查看环境信息：

从github 克隆代码，然后 pip 安装依赖包。

# clone the repository.
git clone https://github.com/SkyworkAI/SkyReels-V2
cd SkyReels-V2
pip install -r requirements.txt

若遇到 flash_attn 一起在安装的问题：可以真的到Releases · Dao-AILab/flash-attention 下载它的编译好的 whl 包，一定要选择对应的版本，

wget https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.5cxx11abiFALSE-cp312-cp312-linux_x86_64.whl
pip install flash_attn-2.7.4.post1+cu12torch2.5cxx11abiFALSE-cp312-cp312-linux_x86_64.whl

然后在执行 pip install -r requirements.txt

四、下载模型文件

目前开源的模型如下：

Type	Model Variant	Recommended Height/Width/Frame	Link
Diffusion Forcing	1.3B-540P	544 * 960 * 97f	🤗Huggingface 🤖 ModelScope
	5B-540P	544 * 960 * 97f	Coming Soon
	5B-720P	720 * 1280 * 121f	Coming Soon
	14B-540P	544 * 960 * 97f	🤗Huggingface 🤖 ModelScope
	14B-720P	720 * 1280 * 121f	Coming Soon
Text-to-Video	1.3B-540P	544 * 960 * 97f	Coming Soon
	5B-540P	544 * 960 * 97f	Coming Soon
	5B-720P	720 * 1280 * 121f	Coming Soon
	14B-540P	544 * 960 * 97f	🤗Huggingface 🤖 ModelScope
	14B-720P	720 * 1280 * 121f	🤗Huggingface 🤖 ModelScope
Image-to-Video	1.3B-540P	544 * 960 * 97f	🤗Huggingface 🤖 ModelScope
	5B-540P	544 * 960 * 97f	Coming Soon
	5B-720P	720 * 1280 * 121f	Coming Soon
	14B-540P	544 * 960 * 97f	🤗Huggingface 🤖 ModelScope
	14B-720P	720 * 1280 * 121f	Coming Soon
Camera Director	5B-540P	544 * 960 * 97f	Coming Soon
	5B-720P	720 * 1280 * 121f	Coming Soon
	14B-720P	720 * 1280 * 121f	Coming Soon

从ModelScope 下载 Diffusion Forcing1.3B-540P 的模型：

git lfs install
git clone https://www.modelscope.cn/Skywork/SkyReels-V2-DF-1.3B-540P.git

五、运行测试

下载完模型后，将下列代码写入：test.sh

model_id=Skywork/SkyReels-V2-DF-1.3B-540P
# synchronous inference
python generate_video_df.py \--model_id ${model_id} \--resolution 540P \--ar_step 0 \--base_num_frames 97 \--num_frames 257 \--overlap_history 17 \--prompt "A graceful white swan with a curved neck and delicate feathers swimming in a serene lake at dawn, its reflection perfectly mirrored in the still water as mist rises from the surface, with the swan occasionally dipping its head into the water to feed." \--addnoise_condition 20 \--offload

运行 sh test.sh 完成后结果保存在：root/SkyReels-V2/result/diffusion_forcing

• 若出现在问题：Error opening output file: File name too long 是由于导出视频的文件名太长所导致的，可以修改 generate_video_df.py 的代码 ：找到下列代码，将 [:100] 修改为 [:10]

    if local_rank == 0:current_time = time.strftime("%Y-%m-%d_%H-%M-%S", time.localtime())video_out_file = f"{args.prompt[:10].replace('/','')}_{args.seed}_{current_time}.mp4"output_path = os.path.join(save_dir, video_out_file)imageio.mimwrite(output_path, video_frames, fps=fps, quality=8, output_params=["-loglevel", "error"])

• 无法使用14B模型，若要使用需要至少一张大于60G显存的显卡

六、使用webui界面

通过 python app.py 运行webui

模型可以放在 SkyReels-V2/Skywork 文件夹下，这个镜像中的模型是使用 ln 创建的链接

# app.py
import os
import gradio as gr
import subprocess
import time
import glob
import torch.distributed as distdef get_model_list(base_dir="./Skywork/"):"""扫描Skywork目录获取可用的模型列表"""try:# 确保目录存在if not os.path.exists(base_dir):os.makedirs(base_dir)return ["请选择模型"]  # 返回默认列表# 扫描目录下的所有子目录model_dirs = glob.glob(os.path.join(base_dir, "*"))model_ids = []for dir_path in model_dirs:if os.path.isdir(dir_path):# 将目录路径转换为model_id格式model_name = os.path.basename(dir_path)model_id = f"{base_dir}/{model_name}"model_ids.append(model_id)# 如果没有找到模型，返回默认列表return model_ids if model_ids else ["请选择模型"]except Exception as e:print(f"扫描模型目录出错: {str(e)}")return ["扫描模型目录出错"]  # 发生错误时返回默认列表def enhance_prompt(prompt):"""使用提示词增强器处理提示词"""try:cmd = ["python","skyreels_v2_infer/pipelines/prompt_enhancer.py","--prompt", str(prompt)]# 执行提示词增强process = subprocess.Popen(cmd,stdout=subprocess.PIPE,stderr=subprocess.STDOUT,text=True,bufsize=1,universal_newlines=True)output = []enhanced_prompt = Nonefor line in process.stdout:print(line, end='')output.append(line)if "Enhanced prompt:" in line:enhanced_prompt = line.split("Enhanced prompt:")[-1].strip()process.wait()if process.returncode != 0:raise RuntimeError("提示词增强失败")return enhanced_prompt if enhanced_prompt else promptexcept Exception as e:print(f"提示词增强失败: {str(e)}")return promptdef run_video_generation(prompt,model_id="选择模型",resolution="540P",ar_step=4,base_num_frames=97,num_frames=257,overlap_history=17,addnoise_condition=20,use_usp=True,offload=True,seed=42,gpu_count=2,image=None,inference_steps=50,fps=24,shift=8.0,guidance_scale=6.0,causal_block_size=5,outdir="./video_out",prompt_enhancer=False,  # 添加prompt_enhancer参数teacache=False  # 添加teacache参数
):"""执行视频生成命令"""# 记录开始时间start_time = time.time()# 确保输出目录存在并获取绝对路径outdir = os.path.abspath(outdir)os.makedirs(outdir, exist_ok=True)# 根据条件处理提示词增强if prompt_enhancer:try:enhanced = enhance_prompt(prompt)if enhanced != prompt:print(f"原始提示词: {prompt}")print(f"增强后的提示词: {enhanced}")prompt = enhancedexcept Exception as e:print(f"提示词增强失败，将使用原始提示词: {str(e)}")# 自动关闭提示词增强功能prompt_enhancer = False# 构建基础命令参数，确保参数顺序正确base_args = ["--prompt", str(prompt),"--model_id", str(model_id),"--resolution", str(resolution),"--ar_step", str(ar_step),"--base_num_frames", str(base_num_frames),"--num_frames", str(num_frames),"--overlap_history", str(overlap_history),"--addnoise_condition", str(addnoise_condition),"--inference_steps", str(inference_steps),"--fps", str(fps),"--shift", str(shift),"--guidance_scale", str(guidance_scale),"--causal_block_size", str(causal_block_size),"--outdir", str(outdir)  # 确保outdir在最后]if teacache:base_args.append("--teacache")if image:base_args.extend(["--image", str(image)])if offload:base_args.append("--offload")if seed != -1:base_args.append("--seed")base_args.append(str(seed))if prompt_enhancer:base_args.append("--prompt_enhancer")# 根据是否使用USP构建完整命令if use_usp:cmd = ["torchrun",f"--nproc_per_node={gpu_count}",f"generate_video_df.py"] + base_args + ["--use_usp"]else:cmd = ["python", "generate_video_df.py"] + base_args# 打印执行命令print("执行命令:", " ".join(cmd))# 确保输出目录存在并获取绝对路径outdir = os.path.abspath(outdir)os.makedirs(outdir, exist_ok=True)# 执行命令并实时显示输出try:process = subprocess.Popen(cmd,stdout=subprocess.PIPE,stderr=subprocess.STDOUT,text=True,bufsize=1,universal_newlines=True)output = []video_path = None# 实时读取并显示输出for line in process.stdout:print(line, end='')  # 实时显示output.append(line)  # 保存输出# 检查是否包含视频路径信息并转换为绝对路径if "Saved video to" in line:video_path = line.split("Saved video to")[-1].strip()video_path = os.path.abspath(video_path)  # 确保是绝对路径print(f"视频保存路径: {video_path}")process.wait()if process.returncode != 0:raise RuntimeError("生成视频失败")# 计算并添加耗时信息end_time = time.time()elapsed_time = end_time - start_timetime_info = f"\n总耗时: {elapsed_time:.2f} 秒"output.append(time_info)return "".join(output), video_pathexcept Exception as e:raise RuntimeError(f"执行命令失败: {str(e)}")def get_available_gpus():"""获取可用的GPU数量和设备ID"""try:import torchgpu_count = torch.cuda.device_count()if gpu_count == 0:return 0, []return gpu_count, list(range(gpu_count))except:return 0, []def scan_video_directory(directory="./video_out"):"""扫描视频目录并返回所有视频文件"""if not os.path.exists(directory):return []video_files = []for ext in ['.mp4', '.avi', '.mov', '.mkv']:files = glob.glob(os.path.join(directory, f'*{ext}'))# 返回(显示名称, 文件路径)的元组for f in files:basename = os.path.basename(f)video_files.append((basename, f))# 按修改时间排序return sorted(video_files, key=lambda x: os.path.getmtime(x[1]), reverse=True)# 在Gradio界面中使用时：
video_list = gr.Dropdown(label="选择视频播放",choices=scan_video_directory(),value=None,interactive=True
)def update_video_list():"""刷新视频列表"""return gr.Dropdown(choices=scan_video_directory())def on_video_select(video_path):"""当选择视频时"""if video_path:  # video_path现在是完整路径return video_pathreturn Nonedef on_generate_click(*args):"""生成视频并更新列表"""try:# 获取参数prompt, model_id, resolution, ar_step, base_num_frames, num_frames, \overlap_history, addnoise_condition, use_usp, offload, seed, gpu_count, \image, inference_steps, fps, shift, guidance_scale, \causal_block_size, outdir, prompt_enhancer, teacache = args  # 添加teacache# 如果ar_step为0(同步模式)，强制设置causal_block_size为1if ar_step == 0:causal_block_size = 1# 如果同时启用了use_usp和prompt_enhancer，先进行提示词增强if use_usp and prompt_enhancer:try:enhanced = enhance_prompt(prompt)if enhanced != prompt:print(f"原始提示词: {prompt}")print(f"增强后的提示词: {enhanced}")prompt = enhancedexcept Exception as e:print(f"提示词增强失败，将使用原始提示词: {str(e)}")# 禁用prompt_enhancer参数，因为提示词已经增强过了prompt_enhancer = False# 执行视频生成output, video_path = run_video_generation(prompt, model_id, resolution, ar_step, base_num_frames, num_frames,overlap_history, addnoise_condition, use_usp, offload, seed, gpu_count,image, inference_steps, fps, shift, guidance_scale,causal_block_size, outdir, prompt_enhancer,teacache)# 更新视频列表new_choices = scan_video_directory()if video_path and os.path.exists(video_path):return output, video_path, gr.Dropdown(choices=new_choices, value=video_path)return output, None, gr.Dropdown(choices=new_choices)except Exception as e:error_msg = f"生成失败: {str(e)}"return error_msg, None, gr.Dropdown(choices=scan_video_directory())# 创建Gradio界面
with gr.Blocks(title="SkyReels-V2 视频生成器") as demo:gr.Markdown("# SkyReels-V2 视频生成器")with gr.Row():with gr.Column():prompt = gr.Textbox(label="提示词 (prompt)",lines=3,value="A graceful white swan with a curved neck and delicate feathers swimming in a serene lake at dawn")# 图片输入image = gr.Image(label="输入图片 (仅用于图像到视频生成)",type="filepath",visible=True)# In the Gradio interface definition, update the model_id dropdown:model_id = gr.Dropdown(choices=get_model_list(),label="模型ID (model_id)",value="选择模型",# Changed to match the available modelallow_custom_value=True)resolution = gr.Dropdown(choices=["540P", "720P"],label="分辨率 (resolution)",value="540P")with gr.Row():num_frames = gr.Number(label="生成帧数 (num_frames)",value=97,precision=0)base_num_frames = gr.Number(label="基础帧数 (base_num_frames)",value=97,precision=0)with gr.Row():inference_steps = gr.Number(label="推理步骤数 (inference_steps)",value=50,precision=0)fps = gr.Number(label="输出视频帧率 (fps)",value=24,precision=0)with gr.Row():shift = gr.Number(label="流匹配参数 (shift)",value=8.0,info="T2V模型推荐8.0，I2V模型推荐5.0")guidance_scale = gr.Number(label="文本匹配强度 (guidance_scale)",value=6.0,info="T2V模型推荐6.0，I2V模型推荐5.0")with gr.Row():ar_step = gr.Number(label="AR步数 (ar_step)",value=4,info="0表示同步模式")causal_block_size = gr.Number(label="因果块大小 (causal_block_size)",value=4,info="异步推理时推荐使用(ar_step>0)")with gr.Row():overlap_history = gr.Number(label="重叠历史帧数 (overlap_history)",value=17)addnoise_condition = gr.Number(label="噪声条件 (addnoise_condition)",value=20)with gr.Row():use_usp = gr.Checkbox(label="启用use_usp(使用多GPU并行推理)",value=True)offload = gr.Checkbox(label="启用offload(使用显存优化)",value=True)prompt_enhancer = gr.Checkbox(label="启用prompt_enhancer(提示词扩展)",value=True,)teacache = gr.Checkbox(label="启用teacache(使用缓存加速)",value=False,)with gr.Row():seed = gr.Number(label="随机种子 (-1为随机)",value=42)gpu_count = gr.Number(label="使用GPU数量",value=2,precision=0)outdir = gr.Textbox(label="输出目录 (outdir)",value="./video_out")# 添加说明gr.Markdown("""### 使用说明1. 单GPU模式：取消勾选"启用USP"选项2. 多GPU模式：- 勾选"启用USP"选项- 设置要使用的GPU数量- 需要固定的随机种子- 多GPU无法使用teacache3. 无法使用14B模型，若要使用需要单张显存大于60G### 更多信息访问 [GitHub仓库](https://github.com/SkyworkAI/SkyReels-V2) 获取详细文档和更新。""")# 生成按钮和输出信息generate_btn = gr.Button("生成视频")output_text = gr.Textbox(label="输出信息", visible=True, lines=3)# 视频预览部分gr.Markdown("### 视频预览")with gr.Row():with gr.Column(scale=4):video_player = gr.Video(label="视频播放",format="mp4",autoplay=False,     # 自动播放选中的视频interactive=True   # 允许用户控制播放)with gr.Column(scale=1):video_list = gr.Dropdown(label="选择视频",choices=scan_video_directory(),value=None,interactive=True)refresh_btn = gr.Button("🔄 刷新视频列表")# 更新事件绑定generate_btn.click(on_generate_click,inputs=[prompt, model_id, resolution, ar_step, base_num_frames, num_frames,overlap_history, addnoise_condition, use_usp, offload, seed, gpu_count,image, inference_steps, fps, shift, guidance_scale,causal_block_size, outdir, prompt_enhancer, teacache],  # 添加teacacheoutputs=[output_text, video_player, video_list])video_list.change(on_video_select,inputs=[video_list],outputs=[video_player])refresh_btn.click(update_video_list,outputs=[video_list])if __name__ == "__main__":try:# 设置CUDA架构列表os.environ["TORCH_CUDA_ARCH_LIST"] = "8.6;8.9"  # 30系和40系的架构# 设置CUDA内存分配策略os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"# 检测GPUgpu_count, gpu_ids = get_available_gpus()# 设置环境变量以消除tokenizers警告os.environ["TOKENIZERS_PARALLELISM"] = "false"if gpu_count == 0:print("警告：未检测到可用的GPU，系统将使用CPU运行（不推荐）")else:print(f"检测到{gpu_count}个GPU {gpu_ids}")os.environ["CUDA_VISIBLE_DEVICES"] = ",".join(map(str, gpu_ids))# 启动Gradio界面，移除不支持的参数demo.launch(server_name="0.0.0.0",server_port=8080,share=False,)finally:# 确保在程序退出时销毁进程组if dist.is_initialized():dist.destroy_process_group()