README_zh.md

---
license: apache-2.0
base_model:
  - THUDM/CogVideoX-5b
language:
  - en
tags:
  - video-generation
  - paddlemix
---

简体中文 | [English](README.md)
# VCtrl
<p style="text-align: center;">
  <p align="center"> 
  <a href="https://huggingface.co/PaddleMIX">🤗 Huggingface Space</a> |
  <a href="https://github.com/PaddlePaddle/PaddleMIX/tree/develop/ppdiffusers/examples/ppvctrl">🌐 Github </a> | 
  <a href="">📜 arxiv </a> |
  <a href="https://pp-vctrl.github.io/">📷 Project </a> 
</p>

## 模型介绍
**VCtrl** 是一个通用的视频生成控制模型，通过引入辅助条件编码器，能够灵活对接各类控制模块，并且在不改变原始生成器的前提下避免了大规模重训练。该模型利用稀疏残差连接实现对控制信号的高效传递，同时通过统一的条件编码流程，将多种控制输入转换为标准化表示，再结合任务特定掩码以提升适应性。得益于这种统一而灵活的设计，VCtrl 可广泛应用于**人物动画**、**场景转换**、**视频编辑**等视频生成场景。下表展示我们在本代提供的视频生成模型列表相关信息:

<table  style="border-collapse: collapse; width: 100%;">
  <tr>
    <th style="text-align: center;">模型名</th>
    <th style="text-align: center;">VCtrl-Canny</th>
    <th style="text-align: center;">VCtrl-Mask</th>
    <th style="text-align: center;">VCtrl-Pose</th>
  </tr>
  <tr>
    <td style="text-align: center;">视频分辨率</td>
    <td colspan="1" style="text-align: center;">720 * 480</td>
    <td colspan="1" style="text-align: center;"> 720*480 </td>
    <td colspan="1" style="text-align: center;"> 720*480 & 480*720 </td>
    </tr>
  <tr>
    <td style="text-align: center;">推理精度</td>
    <td colspan="3" style="text-align: center;"><b>FP16(推荐)</b></td>
  </tr>
  <tr>
    <td style="text-align: center;">单GPU显存消耗</td>
    <td colspan="3"  style="text-align: center;"><b>V100: 32GB minimum*</b></td>
  </tr>
  <tr>
    <td style="text-align: center;">推理速度<br>(Step = 25, FP16)</td>
    <td colspan="3" style="text-align: center;">单卡A100: ~300秒(49帧)<br>单卡V100: ~400秒(49帧)</td>
  </tr>
  <tr>
    <td style="text-align: center;">提示词语言</td>
    <td colspan="5" style="text-align: center;">English*</td>
  </tr>
  <tr>
    <td style="text-align: center;">提示词长度上限</td>
    <td colspan="3" style="text-align: center;">224 Tokens</td>
  </tr>
  <tr>
    <td style="text-align: center;">视频长度</td>
    <td colspan="3" style="text-align: center;">T2V模型只支持49帧，I2V模型可以扩展为任意帧</td>
  </tr>
  <tr>
    <td style="text-align: center;">帧率</td>
    <td colspan="3" style="text-align: center;">30 帧 / 秒 </td>
  </tr>
</table>

##  快速开始 🤗

本模型已经支持使用 paddlemix 的 ppdiffusers 库进行部署，你可以按照以下步骤进行部署。

**我们推荐您进入我们的 [github](https://github.com/PaddlePaddle/PaddleMIX/tree/develop/ppdiffusers/examples/ppvctrl)以获得更好的体验。**

1. 安装对应的依赖

```shell
# 克隆 PaddleMIX 仓库
git clone https://github.com/PaddlePaddle/PaddleMIX.git
#安装paddlemix
cd PaddleMIX
pip install -e .
# 安装ppdiffusers
pip install -e ppdiffusers
# 安装paddlenlp
pip install paddlenlp==v3.0.0-beta2
# 进入 vctrl目录
cd ppdiffusers/examples/ppvctrl
# 安装其他所需的依赖
pip install -r requirements.txt
#安装paddlex
pip install paddlex==3.0.0b2

```

2. 运行代码

```python
import os
import paddle
import numpy as np
from decord import VideoReader
from moviepy.editor import ImageSequenceClip
from PIL import Image
from ppdiffusers import (
    CogVideoXDDIMScheduler,
    CogVideoXTransformer3DVCtrlModel,
    CogVideoXVCtrlPipeline,
    VCtrlModel,
)
def write_mp4(video_path, samples, fps=8):
    clip = ImageSequenceClip(samples, fps=fps)
    clip.write_videofile(video_path, audio_codec="aac")


def save_vid_side_by_side(batch_output, validation_control_images, output_folder, fps):
    flattened_batch_output = [img for sublist in batch_output for img in sublist]
    ori_video_path = output_folder + "/origin_predict.mp4"
    video_path = output_folder + "/test_1.mp4"
    ori_final_images = []
    final_images = []
    outputs = []

    def get_concat_h(im1, im2):
        dst = Image.new("RGB", (im1.width + im2.width, max(im1.height, im2.height)))
        dst.paste(im1, (0, 0))
        dst.paste(im2, (im1.width, 0))
        return dst

    for image_list in zip(validation_control_images, flattened_batch_output):
        predict_img = image_list[1].resize(image_list[0].size)
        result = get_concat_h(image_list[0], predict_img)
        ori_final_images.append(np.array(image_list[1]))
        final_images.append(np.array(result))
        outputs.append(np.array(predict_img))
    write_mp4(ori_video_path, ori_final_images, fps=fps)
    write_mp4(video_path, final_images, fps=fps)
    output_path = output_folder + "/output.mp4"
    write_mp4(output_path, outputs, fps=fps)


def load_images_from_folder_to_pil(folder):
    images = []
    valid_extensions = {".jpg", ".jpeg", ".png", ".bmp", ".gif", ".tiff"}

    def frame_number(filename):
        new_pattern_match = re.search("frame_(\\d+)_7fps", filename)
        if new_pattern_match:
            return int(new_pattern_match.group(1))
        matches = re.findall("\\d+", filename)
        if matches:
            if matches[-1] == "0000" and len(matches) > 1:
                return int(matches[-2])
            return int(matches[-1])
        return float("inf")

    sorted_files = sorted(os.listdir(folder), key=frame_number)
    for filename in sorted_files:
        ext = os.path.splitext(filename)[1].lower()
        if ext in valid_extensions:
            img = Image.open(os.path.join(folder, filename)).convert("RGB")
            images.append(img)
    return images


def load_images_from_video_to_pil(video_path):
    images = []
    vr = VideoReader(video_path)
    length = len(vr)
    for idx in range(length):
        frame = vr[idx].asnumpy()
        images.append(Image.fromarray(frame))
    return images


validation_control_images = load_images_from_video_to_pil('your_path')
prompt = 'Group of fishes swimming in aquarium.'
vctrl = VCtrlModel.from_pretrained(
            paddlemix/vctrl-5b-t2v-canny,
            low_cpu_mem_usage=True,
            paddle_dtype=paddle.float16
        )
pipeline = CogVideoXVCtrlPipeline.from_pretrained(
            paddlemix/cogvideox-5b-vctrl, 
            vctrl=vctrl, 
            paddle_dtype=paddle.float16, 
            low_cpu_mem_usage=True,
            map_location="cpu",
        )
pipeline.scheduler = CogVideoXDDIMScheduler.from_config(pipeline.scheduler.config, timestep_spacing="trailing")
pipeline.vae.enable_tiling()
pipeline.vae.enable_slicing()
task='canny'
final_result=[]
video = pipeline(
        prompt=prompt,
        num_inference_steps=25,
        num_frames=49,
        guidance_scale=35,
        generator=paddle.Generator().manual_seed(42),
        conditioning_frames=validation_control_images[:num_frames],
        conditioning_frame_indices=list(range(num_frames)),
        conditioning_scale=1.0,
        width=720,
        height=480,
        task='canny',
        conditioning_masks=validation_mask_images[:num_frames] if task == "mask" else None,
        vctrl_layout_type='spacing',
    ).frames[0]
final_result.append(video)
save_vid_side_by_side(final_result, validation_control_images[:num_frames], 'save.mp4', fps=30)
```

## 深入研究

欢迎进入我们的 [github]("https://github.com/PaddlePaddle/PaddleMIX/tree/develop/ppdiffusers/examples/ppvctrl")，你将获得：

1. 更加详细的技术细节介绍和代码解释。
2. 控制条件的提取算法细节。
3. 模型推理的详细代码。
4. 项目更新日志动态，更多互动机会。
5. PaddleMix工具链，帮助您更好的使用模型。

<!-- ## 引用

```
@article{yang2024cogvideox,
  title={VCtrl: Enabling Versatile Controls for Video Diffusion Models},
  year={2025}
}
``` -->