AnimatedDiff ControlNet SDXL Example

This document provides a step-by-step guide to setting up and running the animatediff_controlnet_sdxl.py script from the Hugging Face repository. The script leverages the diffusers-sdxl-controlnet library to generate animated images using ControlNet and SDXL models.

Prerequisites

Before running the script, ensure you have the necessary dependencies installed. You can install them using the following commands:

System Dependencies

sudo apt-get update && sudo apt-get install git-lfs cbm ffmpeg

Python Dependencies

pip install git+https://huggingface.co/svjack/diffusers-sdxl-controlnet
pip install transformers peft sentencepiece moviepy==1.0.3 controlnet_aux

Clone the Repository

git clone https://huggingface.co/svjack/diffusers-sdxl-controlnet
cp diffusers-sdxl-controlnet/girl-pose.gif .
cp diffusers-sdxl-controlnet/girl_beach.mp4 .

Script Modifications

The script requires some modifications to work correctly. Specifically, you need to comment out certain lines related to LoRA processors:

'''
drop     #LoRAAttnProcessor2_0,
    #LoRAXFormersAttnProcessor,
'''

GIF to Frames Conversion

The script includes a function to convert a GIF into individual frames. This is useful for preparing input data for the animation pipeline.

from PIL import Image, ImageSequence
import os

def gif_to_frames(gif_path, output_folder):
    # Open the GIF file
    gif = Image.open(gif_path)
    
    # Ensure the output folder exists
    if not os.path.exists(output_folder):
        os.makedirs(output_folder)
    
    # Iterate through each frame of the GIF
    for i, frame in enumerate(ImageSequence.Iterator(gif)):
        # Copy the frame
        frame_copy = frame.copy()
        
        # Save the frame to the specified folder
        frame_path = os.path.join(output_folder, f"frame_{i:04d}.png")
        frame_copy.save(frame_path)
    
    print(f"Successfully extracted {i + 1} frames to {output_folder}")

# Example call
gif_to_frames("girl-pose.gif", "girl_pose_frames")

Use this girl pose as pose source video (gif)

Running the Script

To run the script, follow these steps:

Add the Script Path to System Path:

import sys
sys.path.insert(0, "diffusers-sdxl-controlnet/examples/community/")
from animatediff_controlnet_sdxl import *
from controlnet_aux.processor import Processor

Load Necessary Libraries and Models:

import torch
from diffusers.models import MotionAdapter
from diffusers import DDIMScheduler
from diffusers.utils import export_to_gif
from diffusers import AutoPipelineForText2Image, ControlNetModel
from diffusers.utils import load_image
from PIL import Image

Load the MotionAdapter Model:

adapter = MotionAdapter.from_pretrained(
    "a-r-r-o-w/animatediff-motion-adapter-sdxl-beta", 
    torch_dtype=torch.float16
)

Configure the Scheduler and ControlNet:

model_id = "svjack/GenshinImpact_XL_Base"
scheduler = DDIMScheduler.from_pretrained(
    model_id,
    subfolder="scheduler",
    clip_sample=False,
    timestep_spacing="linspace",
    beta_schedule="linear",
    steps_offset=1,
)

controlnet = ControlNetModel.from_pretrained(
    "thibaud/controlnet-openpose-sdxl-1.0",
    torch_dtype=torch.float16,
).to("cuda")

Load the AnimateDiffSDXLControlnetPipeline:

pipe = AnimateDiffSDXLControlnetPipeline.from_pretrained(
    model_id,
    controlnet=controlnet,
    motion_adapter=adapter,
    scheduler=scheduler,
    torch_dtype=torch.float16,
).to("cuda")

Enable Memory Saving Features:

pipe.enable_vae_slicing()
pipe.enable_vae_tiling()

Load Conditioning Frames:

import os
folder_path = "girl_pose_frames/"
frames = os.listdir(folder_path)
frames = list(filter(lambda x: x.endswith(".png"), frames))
frames.sort()
conditioning_frames = list(map(lambda x: Image.open(os.path.join(folder_path ,x)).resize((1024, 1024)), frames))[:16]

Process Conditioning Frames:

p2 = Processor("openpose")
cn2 = [p2(frame) for frame in conditioning_frames]

Define Prompts:

prompt = '''
solo,Xiangling\(genshin impact\),1girl,
full body professional photograph of a stunning detailed, sharp focus, dramatic 
cinematic lighting, octane render  unreal engine (film grain, blurry background
'''
prompt = "solo,Xiangling\(genshin impact\),1girl,full body professional photograph of a stunning detailed"
negative_prompt = "bad quality, worst quality, jpeg artifacts, ugly"

Generate Output: (Use Genshin Impact character Xiangling)

prompt = '''
solo,Xiangling\(genshin impact\),1girl,
full body professional photograph of a stunning detailed, sharp focus, dramatic 
cinematic lighting, octane render  unreal engine (film grain, blurry background
'''
prompt = "solo,Xiangling\(genshin impact\),1girl,full body professional photograph of a stunning detailed"

#prompt = "solo,Xiangling\(genshin impact\),1girl"
negative_prompt = "bad quality, worst quality, jpeg artifacts, ugly"

generator = torch.Generator(device="cpu").manual_seed(0)
output = pipe(
    prompt=prompt,
    negative_prompt=negative_prompt,
    num_inference_steps=50,
    guidance_scale=20,
    controlnet_conditioning_scale = 1.0,
    width=512,
    height=768,
    num_frames=16,
    conditioning_frames=cn2,
    generator = generator
)

Export Frames to GIF:

frames = output.frames[0]
export_to_gif(frames, "xiangling_animation.gif")

Display the Result:

from IPython import display
display.Image("xiangling_animation.gif")

Target gif

Use Anime Upscale in https://github.com/svjack/APISR

Run in Command line

animatediff_controlnet_sdxl_run_script.py

import sys
sys.path.insert(0, "diffusers-sdxl-controlnet/examples/community/")
from animatediff_controlnet_sdxl import *

import argparse
from moviepy.editor import VideoFileClip, ImageSequenceClip
import os
import torch
from diffusers.models import MotionAdapter
from diffusers import DDIMScheduler, AutoPipelineForText2Image, ControlNetModel
from diffusers.utils import export_to_gif
from PIL import Image
from controlnet_aux.processor import Processor

# 初始化 MotionAdapter 和 ControlNetModel
adapter = MotionAdapter.from_pretrained("a-r-r-o-w/animatediff-motion-adapter-sdxl-beta", torch_dtype=torch.float16)

def initialize_pipeline(model_id):
    scheduler = DDIMScheduler.from_pretrained(model_id, subfolder="scheduler", clip_sample=False, timestep_spacing="linspace", beta_schedule="linear", steps_offset=1)
    controlnet = ControlNetModel.from_pretrained("thibaud/controlnet-openpose-sdxl-1.0", torch_dtype=torch.float16).to("cuda")

    # 初始化 AnimateDiffSDXLControlnetPipeline
    pipe = AnimateDiffSDXLControlnetPipeline.from_pretrained(
        model_id,
        controlnet=controlnet,
        motion_adapter=adapter,
        scheduler=scheduler,
        torch_dtype=torch.float16,
    ).to("cuda")
    pipe.enable_vae_slicing()
    pipe.enable_vae_tiling()
    return pipe

def split_video_into_frames(input_video_path, num_frames, temp_folder='temp_frames'):
    """
    将视频处理成指定帧数的视频，并保持原始的帧率。

    :param input_video_path: 输入视频文件路径
    :param num_frames: 目标帧数
    :param temp_folder: 临时文件夹路径
    """
    clip = VideoFileClip(input_video_path)
    original_duration = clip.duration
    segment_duration = original_duration / num_frames

    if not os.path.exists(temp_folder):
        os.makedirs(temp_folder)

    for i in range(num_frames):
        frame_time = i * segment_duration
        frame_path = os.path.join(temp_folder, f'frame_{i:04d}.png')
        clip.save_frame(frame_path, t=frame_time)

    frame_paths = [os.path.join(temp_folder, f'frame_{i:04d}.png') for i in range(num_frames)]
    final_clip = ImageSequenceClip(frame_paths, fps=clip.fps)
    final_clip.write_videofile("resampled_video.mp4", codec='libx264')

    print(f"新的视频已保存到 resampled_video.mp4，包含 {num_frames} 个帧，并保持原始的帧率。")

def generate_video_with_prompt(input_video_path, prompt, model_id, gif_output_path, seed=0, num_frames=16, keep_imgs=False, temp_folder='temp_frames', num_inference_steps=50, guidance_scale=20, controlnet_conditioning_scale=1.0, width=512, height=768):
    """
    生成带有文本提示的视频。

    :param input_video_path: 输入视频文件路径
    :param prompt: 文本提示
    :param model_id: 模型ID
    :param gif_output_path: GIF 输出文件路径
    :param seed: 随机种子
    :param num_frames: 目标帧数
    :param keep_imgs: 是否保留临时图片
    :param temp_folder: 临时文件夹路径
    :param num_inference_steps: 推理步数
    :param guidance_scale: 引导比例
    :param controlnet_conditioning_scale: ControlNet 条件比例
    :param width: 输出宽度
    :param height: 输出高度
    """
    split_video_into_frames(input_video_path, num_frames, temp_folder)

    folder_path = temp_folder
    frames = os.listdir(folder_path)
    frames = list(filter(lambda x: x.endswith(".png"), frames))
    frames.sort()
    conditioning_frames = list(map(lambda x: Image.open(os.path.join(folder_path, x)).resize((1024, 1024)), frames))[:num_frames]

    p2 = Processor("openpose")
    cn2 = [p2(frame) for frame in conditioning_frames]

    negative_prompt = "bad quality, worst quality, jpeg artifacts, ugly"
    generator = torch.Generator(device="cuda").manual_seed(seed)

    pipe = initialize_pipeline(model_id)

    output = pipe(
        prompt=prompt,
        negative_prompt=negative_prompt,
        num_inference_steps=num_inference_steps,
        guidance_scale=guidance_scale,
        controlnet_conditioning_scale=controlnet_conditioning_scale,
        width=width,
        height=height,
        num_frames=num_frames,
        conditioning_frames=cn2,
        generator=generator
    )

    frames = output.frames[0]
    export_to_gif(frames, gif_output_path)

    print(f"生成的 GIF 已保存到 {gif_output_path}")

    if not keep_imgs:
        # 删除临时文件夹
        import shutil
        shutil.rmtree(temp_folder)

if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="生成带有文本提示的视频")
    parser.add_argument("input_video", help="输入视频文件路径")
    parser.add_argument("prompt", help="文本提示")
    parser.add_argument("model_id", help="模型ID")
    parser.add_argument("gif_output_path", help="GIF 输出文件路径")
    parser.add_argument("--seed", type=int, default=0, help="随机种子")
    parser.add_argument("--num_frames", type=int, default=16, help="目标帧数")
    parser.add_argument("--keep_imgs", action="store_true", help="是否保留临时图片")
    parser.add_argument("--temp_folder", default='temp_frames', help="临时文件夹路径")
    parser.add_argument("--num_inference_steps", type=int, default=50, help="推理步数")
    parser.add_argument("--guidance_scale", type=float, default=20.0, help="引导比例")
    parser.add_argument("--controlnet_conditioning_scale", type=float, default=1.0, help="ControlNet 条件比例")
    parser.add_argument("--width", type=int, default=512, help="输出宽度")
    parser.add_argument("--height", type=int, default=768, help="输出高度")

    args = parser.parse_args()

    generate_video_with_prompt(args.input_video, args.prompt, args.model_id, args.gif_output_path, args.seed, args.num_frames,
      args.keep_imgs, args.temp_folder, args.num_inference_steps, args.guidance_scale, args.controlnet_conditioning_scale, args.width, args.height)

python animatediff_controlnet_sdxl_run_script.py girl_beach.mp4 \
 "solo,Xiangling\(genshin impact\),1girl,full body professional photograph of a stunning detailed, drink tea use chinese cup" \
  "svjack/GenshinImpact_XL_Base" \
   xiangling_tea_animation.gif --num_frames 16 --temp_folder temp_frames

Pose: girl_beach.mp4
Output: xiangling_tea_animation.gif
Upscaled:

Some Other Samples

Makise Kurisu in Steins Gate

python animatediff_controlnet_sdxl_run_script.py girl_beach.mp4 \
 "1girl, Makise Kurisu, masterpiece, white lab coat, red tie, laboratory" \
  "cagliostrolab/animagine-xl-3.1" \
   Makise_Kurisu_animation_short.gif --num_frames 16 --temp_folder temp_frames --guidance_scale 20 --controlnet_conditioning_scale 0.3

Output: Makise_Kurisu_animation_short.gif
Upscaled:

Souryuu Asuka Langley in EVA

python animatediff_controlnet_sdxl_run_script.py girl_beach.mp4 \
 "1girl, souryuu asuka langley, masterpiece" \
  "cagliostrolab/animagine-xl-3.1" \
   asuka_langley_animation_short.gif --num_frames 16 --temp_folder temp_frames --guidance_scale 20 --controlnet_conditioning_scale 0.3 --num_inference_steps 50

Output: asuka_langley_animation_short.gif
Upscaled:

python animatediff_controlnet_sdxl_run_script.py girl_beach.mp4 \
 "1girl, souryuu asuka langley, masterpiece, neon genesis evangelion, solo, upper body, v, smile, looking at viewer, outdoors, night" \
  "cagliostrolab/animagine-xl-3.1" \
   asuka_langley_animation_long.gif --num_frames 16 --temp_folder temp_frames --guidance_scale 20 --controlnet_conditioning_scale 0.3

Output: asuka_langley_animation_long.gif
Upscaled:

XiangLing in Genshin Impact

produce_gif_script.py

python produce_gif_script.py xiangling_video_seed.csv "svjack/GenshinImpact_XL_Base" xiangling_gif_dir \
 --num_frames 16 --temp_folder temp_frames --seed 0 --controlnet_conditioning_scale 0.3

Conclusion

This script demonstrates how to use the diffusers-sdxl-controlnet library to generate animated images with ControlNet and SDXL models. By following the steps outlined above, you can create and visualize your own animated sequences.