aniportrait / README.md

Update README.md

d077c5b verified 6 months ago

27.7 kB

	---
	license: apache-2.0
	---

	This repository contains a pruned and partially reorganized version of [AniPortrait](https://github.com/Zejun-Yang/AniPortrait), with some new features.

	```
	@misc{wei2024aniportrait,
	title={AniPortrait: Audio-Driven Synthesis of Photorealistic Portrait Animations},
	author={Huawei Wei and Zejun Yang and Zhisheng Wang},
	year={2024},
	eprint={2403.17694},
	archivePrefix={arXiv},
	primaryClass={cs.CV}
	}
	```

	# Added Features

	## CPU Offloading

	In order to maximize VRAM, CPU offloading is made available and enabled by default when using the command-line.

	- In python, use `pipeline.enable_model_cpu_offload(gpu_id: int=0)` to enable it.
	- In the command-line, pass `--no-offload` or `-no` to disable it.

	## Video Reference

	In addition to using an image as reference, this repository permits using a video as a reference image using a modified temporally-aware reference U-Net.

	### Automatic Masked Composition

	When using a video reference, you can pass `paste_back=True` or `--paste-back/-pb` to automatically re-composite the generated face onto the reference video.

	### Automatic Face Cropping

	In addition to the above, you can pass `crop_to_face=True` or `--crop-to-face/-ctf` to automatically crop the input video to the detected face region prior to inference, then un-crop and compose the results. This means you can directly use this feature on HD video without needing to pre-process any of the inputs.

	### Examples

	The following examples were produced using the above-referenced procedure, then interpolated and post-processed using [enfugue](https://github.com/painebenjamin/app.enfugue.ai).

	They were made using the source videos at Full HD, using the following command-line format:

	```sh
	aniportrait video.mp4 --video video.mp4 --audio audio.mp3 --crop-to-face --num-inference-steps 50 -cfg 4.5
	```

	All videos sourced from [Pexels](https://www.pexels.com/), voices are AI-generated reading [Attention is All You Need (ARXIV:1706.03762)](https://arxiv.org/abs/1706.03762)

	<table>
	<tr>
	<td>
	<video controls src="https://cdn-uploads.huggingface.co/production/uploads/64429aaf7feb866811b12f73/nW0kiagVIP2Nr_K_SjJUP.mp4"</video>
	</td>
	<td>
	<video controls src="https://cdn-uploads.huggingface.co/production/uploads/64429aaf7feb866811b12f73/FOj4drfm2OLa-n7JxYyIj.mp4"></video>
	</td>
	<td>
	<video controls src="https://cdn-uploads.huggingface.co/production/uploads/64429aaf7feb866811b12f73/Tu2ac8Cuvaexo_S-kgpPv.mp4"></video>
	</td>
	</tr>
	</table>

	# Usage

	## Installation

	First, install the AniPortrait package into your python environment. If you're creating a new environment for AniPortrait, be sure you also specify the version of torch you want with CUDA support, or else this will try to run only on CPU.

	```sh
	pip install git+https://github.com/painebenjamin/aniportrait.git
	```

	## Command-Line

	A command-line utility `aniportrait` is installed with the package.

	```sh
	Usage: aniportrait [OPTIONS] INPUT_IMAGE_OR_VIDEO

	Run AniPortrait on an input image with a video, and/or audio file. When only
	a video file is provided, a video-to-video (face reenactment) animation is
	performed. When only an audio file is provided, an audio-to-video (lip-sync)
	animation is performed. When both a video and audio file are provided, a
	video-to-video animation is performed with the audio as guidance for the
	face and mouth movements.

	Options:
	-v, --video FILE Video file to drive the animation.
	-a, --audio FILE Audio file to drive the animation.
	-fps, --frame-rate INTEGER Video FPS. Also controls the sampling rate
	of the audio. Will default to the video FPS
	if a video file is provided, or 30 if not.
	-cfg, --guidance-scale FLOAT Guidance scale for the diffusion process.
	[default: 3.5]
	-ns, --num-inference-steps INTEGER
	Number of diffusion steps. [default: 20]
	-cf, --context-frames INTEGER Number of context frames to use. [default:
	16]
	-co, --context-overlap INTEGER Number of context frames to overlap.
	[default: 4]
	-nf, --num-frames INTEGER An explicit number of frames to use. When
	not passed, use the length of the audio or
	video
	-s, --seed INTEGER Random seed.
	-w, --width INTEGER Output video width. Defaults to the input
	image width.
	-h, --height INTEGER Output video height. Defaults to the input
	image height.
	-m, --model TEXT HuggingFace model name.
	-nh, --no-half Do not use half precision.
	-no, --no-offload Do not offload to the CPU to preserve GPU
	memory.
	-g, --gpu-id INTEGER GPU ID to use.
	-sf, --model-single-file Download and use a single file instead of a
	directory.
	-cf, --config-file TEXT Config file to use when using the model-
	single-file option. Accepts a path or a
	filename in the same directory as the single
	file. Will download from the repository
	passed in the model option if not provided.
	[default: config.json]
	-mf, --model-filename TEXT The model file to download when using the
	model-single-file option. [default:
	aniportrait.safetensors]
	-rs, --remote-subfolder TEXT Remote subfolder to download from when using
	the model-single-file option.
	-cd, --cache-dir DIRECTORY Cache directory to download to. Default uses
	the huggingface cache.
	-o, --output FILE Output file. [default: output.mp4]
	-pb, --paste-back Paste the original background back in.
	-pbcf, --paste-back-color-fix [adain\|wavelet]
	Color fix method to use when pasting back.
	[default: wavelet]
	-ctf, --crop-to-face Crop the input to the face prior to
	execution, then merge the cropped result
	with the uncropped image. Implies --paste-
	back.
	-pop, --pose-output FILE When passed, save the pose image(s) to this
	file.
	-mop, --mask-output FILE When passed, save the mask image(s) to this
	file.
	-cop, --combined-output FILE When passed, save the combined image(s) to
	this file.
	-mb, --mask-blur INTEGER Amount of blur to apply to the mask when
	using cropping or pasting. [default: 15]
	-md, --mask-dilate INTEGER Amount of dilation to apply to the mask when
	using cropping or pasting. [default: 31]
	-ms, --mask-slow Use a slower, more accurate mask generation
	method.
	-lss, --leading-seconds-silence FLOAT
	Seconds of silence to add to the beginning
	of the audio. [default: 0.0]
	-tss, --trailing-seconds-silence FLOAT
	Seconds of silence to add to the end of the
	audio. [default: 0.0]
	--help Show this message and exit.
	```

	## Python

	You can create the pipeline, automatically pulling the weights from this repository, either as individual models:

	```py
	from aniportrait import AniPortraitPipeline
	pipeline = AniPortraitPipeline.from_pretrained(
	"benjamin-paine/aniportrait",
	torch_dtype=torch.float16,
	variant="fp16",
	).to("cuda", dtype=torch.float16)
	```

	Or, as a single file:

	```py
	from aniportrait import AniPortraitPipeline
	pipeline = AniPortraitPipeline.from_single_file(
	"benjamin-paine/aniportrait",
	torch_dtype=torch.float16,
	variant="fp16",
	).to("cuda", dtype=torch.float16)
	```

	The `AniPortraitPipeline` is a mega pipeline, capable of instantiating and executing other pipelines. It provides the following functions:

	## Workflows

	### img2img

	```py
	pipeline.img2img(
	reference_image: PIL.Image.Image,
	pose_reference_image: PIL.Image.Image,
	num_inference_steps: int,
	guidance_scale: float,
	eta: float=0.0,
	reference_pose_image: Optional[Image.Image]=None,
	generation: Optional[Union[torch.Generator, List[torch.Generator]]]=None,
	output_type: Optional[str]="pil",
	return_dict: bool=True,
	callback: Optional[Callable[[int, int, torch.FloatTensor], None]]=None,
	callback_steps: Optional[int]=None,
	width: Optional[int]=None,
	height: Optional[int]=None,
	**kwargs: Any
	) -> Pose2VideoPipelineOutput
	```

	Using a reference image (for structure) and a pose reference image (for pose), render an image of the former in the pose of the latter.
	- The pose reference image here is an unprocessed image, from which the face pose will be extracted.
	- Optionally pass `reference_pose_image` to designate the pose of `reference_image`. When not passed, the pose of `reference_image` is automatically detected.

	### vid2vid

	```py
	pipeline.vid2vid(
	reference_image: PIL.Image.Image,
	pose_reference_images: List[PIL.Image.Image],
	num_inference_steps: int,
	guidance_scale: float,
	eta: float=0.0,
	reference_pose_image: Optional[Image.Image]=None,
	generation: Optional[Union[torch.Generator, List[torch.Generator]]]=None,
	output_type: Optional[str]="pil",
	return_dict: bool=True,
	callback: Optional[Callable[[int, int, torch.FloatTensor], None]]=None,
	callback_steps: Optional[int]=None,
	width: Optional[int]=None,
	height: Optional[int]=None,
	video_length: Optional[int]=None,
	context_schedule: str="uniform",
	context_frames: int=16,
	context_overlap: int=4,
	context_batch_size: int=1,
	interpolation_factor: int=1,
	use_long_video: bool=True,
	**kwargs: Any
	) -> Pose2VideoPipelineOutput
	```

	Using a reference image (for structure) and a sequence of pose reference images (for pose), render a video of the former in the poses of the latter, using context windowing for long-video generation when the poses are longer than 16 frames.
	- Optionally pass `use_long_video = false` to disable using the long video pipeline.
	- Optionally pass `reference_pose_image` to designate the pose of `reference_image`. When not passed, the pose of `reference_image` is automatically detected.
	- Optionally pass `video_length` to use this many frames. Default is the same as the length of the pose reference images.

	### audio2vid

	```py
	pipeline.audio2vid(
	audio: str,
	reference_image: PIL.Image.Image,
	num_inference_steps: int,
	guidance_scale: float,
	fps: int=30,
	eta: float=0.0,
	reference_pose_image: Optional[Image.Image]=None,
	pose_reference_images: Optional[List[PIL.Image.Image]]=None,
	generation: Optional[Union[torch.Generator, List[torch.Generator]]]=None,
	output_type: Optional[str]="pil",
	return_dict: bool=True,
	callback: Optional[Callable[[int, int, torch.FloatTensor], None]]=None,
	callback_steps: Optional[int]=None,
	width: Optional[int]=None,
	height: Optional[int]=None,
	video_length: Optional[int]=None,
	context_schedule: str="uniform",
	context_frames: int=16,
	context_overlap: int=4,
	context_batch_size: int=1,
	interpolation_factor: int=1,
	use_long_video: bool=True,
	pose_filename: Optional[str]=None,
	leading_seconds_silence: float=0.0,
	trailing_seconds_silence: float=0.0,
	**kwargs: Any
	) -> Pose2VideoPipelineOutput
	```

	Using an audio file, draw `fps` face pose images per second for the duration of the audio. Then, using those face pose images, render a video.
	- Optionally include a list of images to extract the poses from prior to merging with audio-generated poses (in essence, pass a video here to control non-speech motion). The default is a moderately active loop of head movement.
	- Optionally pass width/height to modify the size. Defaults to reference image size.
	- Optionally pass `use_long_video = false` to disable using the long video pipeline.
	- Optionally pass `reference_pose_image` to designate the pose of `reference_image`. When not passed, the pose of `reference_image` is automatically detected.
	- Optionally pass `video_length` to use this many frames. Default is the same as the length of the pose reference images or the length of the audio frames (when translated to the proper FPS) - whichever is shorter.
	- Optionally pass `leading_seconds_silence` and/or `trailing_seconds_silence` to add silent frame(s) to the beginning and/or end of the audio. This will be adjusted for your passed or detected frame rate.
	- Optionally pass `pose_filename`, `mask_filename` and/or `combined_filename` to save the pose, mask and/or combined frames to this video file for debugging.

	### audiovid2vid

	```py
	pipeline.audiovid2vid(
	audio: str,
	reference_image: List[Image.Image],
	num_inference_steps: int=25,
	guidance_scale: float=3.5,
	fps: int=30,
	eta: float=0.0,
	reference_pose_image: Optional[Image.Image]=None,
	pose_reference_images: Optional[List[Image.Image]]=None,
	generation: Optional[Union[torch.Generator, List[torch.Generator]]]=None,
	output_type: Optional[str]="pil",
	return_dict: bool=True,
	callback: Optional[Callable[[int, int, torch.FloatTensor], None]]=None,
	callback_steps: Optional[int]=None,
	context_schedule: str="uniform",
	context_frames: int=16,
	context_overlap: int=4,
	context_batch_size: int=1,
	interpolation_factor: int=1,
	width: Optional[int]=None,
	height: Optional[int]=None,
	video_length: Optional[int]=None,
	use_long_video: bool=True,
	paste_back: bool=True,
	paste_back_color_fix: Optional[Literal["wavelet", "adain"]]="wavelet",
	crop_to_face: bool=False,
	crop_to_face_target_size: Optional[int]=512,
	crop_to_face_padding: Optional[int]=64,
	mask_filename: Optional[str]=None,
	pose_filename: Optional[str]=None,
	combined_filename: Optional[str]=None,
	mask_dilate: Optional[int]=31,
	mask_gaussian_kernel_size: Optional[int]=15,
	mask_first_frame: bool=True,
	leading_seconds_silence: float=0.0,
	trailing_seconds_silence: float=0.0,
	**kwargs: Any
	) -> Pose2VideoPipelineOutput:
	```

	Using an audio file, draw `fps` face pose images per second for the duration of the audio. Then, using those face pose images, render a video using a video as a reference.
	- Optionally pass width/height to modify the size. Defaults to reference image size.
	- Optionally pass `use_long_video = false` to disable using the long video pipeline.
	- Optionally pass `video_length` to use this many frames. Default is the same as the length of the pose reference images or the length of the audio frames (when translated to the proper FPS) - whichever is shorter.
	- Optionally pass `paste_back = true` to re-composite the output onto the input.
	- When using `paste_back`, the face is color-fixed when re-pasting in order to reduce visible difference. The default method is `wavelet`, pass `adain` or `None` for other options.
	- Optionally pass `crop_to_face = true` to crop all images to the face region (with padding) prior to diffusion. This implies `paste_back = true`.
	- When using `crop_to_face`, we must first identify where the faces are in a potentially large image - to do this we perform a slow tiled face detection across the whole image. In order to reduce time to generate additional masks, the default behavior is to only do this once, then use the faces from the first frame to know where to detect in subsequent frames. Set this to `false` to perform tiled faced detection on every frame of the input - this is slower but will allow for more variability in the frame.
	- Optionally pass `leading_seconds_silence` and/or `trailing_seconds_silence` to add silent frame(s) to the beginning and/or end of the audio. This will be adjusted for your passed or detected frame rate.
	- Optionally pass `pose_filename`, `mask_filename` and/or `combined_filename` to save the pose, mask and/or combined frames to this video file for debugging.

	## Internals/Helpers

	### img2pose

	```py
	pipeline.img2pose(
	reference_image: PIL.Image.Image,
	width: Optional[int]=None,
	height: Optional[int]=None
	) -> PIL.Image.Image
	```

	Detects face landmarks in an image and draws a face pose image.
	- Optionally modify the original width and height.

	### vid2pose

	```py
	pipeline.vid2pose(
	reference_image: PIL.Image.Image,
	retarget_image: Optional[PIL.Image.Image],
	width: Optional[int]=None,
	height: Optional[int]=None
	) -> List[PIL.Image.Image]
	```

	Detects face landmarks in a series of images and draws pose images.
	- Optionally modify the original width and height.
	- Optionally retarget to a different face position, useful for video-to-video tasks.

	### audio2pose

	```py
	pipeline.audio2pose(
	audio_path: str,
	fps: int=30,
	reference_image: Optional[PIL.Image.Image]=None,
	pose_reference_images: Optional[List[PIL.Image.Image]]=None,
	width: Optional[int]=None,
	height: Optional[int]=None
	) -> List[PIL.Image.Image]
	```

	Using an audio file, draw `fps` face pose images per second for the duration of the audio.
	- Optionally include a reference image to extract the face shape and initial position from. Default has a generic androgynous face shape.
	- Optionally include a list of images to extract the poses from prior to merging with audio-generated poses (in essence, pass a video here to control non-speech motion). The default is a moderately active loop of head movement.
	- Optionally pass width/height to modify the size. Defaults to reference image size, then pose image sizes, then 256.

	### pose2img

	```py
	pipeline.pose2img(
	reference_image: PIL.Image.Image,
	pose_image: PIL.Image.Image,
	num_inference_steps: int,
	guidance_scale: float,
	eta: float=0.0,
	reference_pose_image: Optional[Image.Image]=None,
	generation: Optional[Union[torch.Generator, List[torch.Generator]]]=None,
	output_type: Optional[str]="pil",
	return_dict: bool=True,
	callback: Optional[Callable[[int, int, torch.FloatTensor], None]]=None,
	callback_steps: Optional[int]=None,
	width: Optional[int]=None,
	height: Optional[int]=None,
	**kwargs: Any
	) -> Pose2VideoPipelineOutput
	```

	Using a reference image (for structure) and a pose image (for pose), render an image of the former in the pose of the latter.
	- The pose image here is a processed face pose. To pass a non-processed face pose, see `img2img`.
	- Optionally pass `reference_pose_image` to designate the pose of `reference_image`. When not passed, the pose of `reference_image` is automatically detected.

	### pose2vid

	```py
	pipeline.pose2vid(
	reference_image: PIL.Image.Image,
	pose_images: List[PIL.Image.Image],
	num_inference_steps: int,
	guidance_scale: float,
	eta: float=0.0,
	reference_pose_image: Optional[Image.Image]=None,
	generation: Optional[Union[torch.Generator, List[torch.Generator]]]=None,
	output_type: Optional[str]="pil",
	return_dict: bool=True,
	callback: Optional[Callable[[int, int, torch.FloatTensor], None]]=None,
	callback_steps: Optional[int]=None,
	width: Optional[int]=None,
	height: Optional[int]=None,
	video_length: Optional[int]=None,
	**kwargs: Any
	) -> Pose2VideoPipelineOutput
	```

	Using a reference image (for structure) and pose images (for pose), render a video of the former in the poses of the latter.
	- The pose images here are a processed face poses. To non-processed face poses, see `vid2vid`.
	- Optionally pass `reference_pose_image` to designate the pose of `reference_image`. When not passed, the pose of `reference_image` is automatically detected.
	- Optionally pass `video_length` to use this many frames. Default is the same as the length of the pose images.

	### pose2vid_long

	```py
	pipeline.pose2vid_long(
	reference_image: PIL.Image.Image,
	pose_images: List[PIL.Image.Image],
	num_inference_steps: int,
	guidance_scale: float,
	eta: float=0.0,
	reference_pose_image: Optional[Image.Image]=None,
	generation: Optional[Union[torch.Generator, List[torch.Generator]]]=None,
	output_type: Optional[str]="pil",
	return_dict: bool=True,
	callback: Optional[Callable[[int, int, torch.FloatTensor], None]]=None,
	callback_steps: Optional[int]=None,
	width: Optional[int]=None,
	height: Optional[int]=None,
	video_length: Optional[int]=None,
	context_schedule: str="uniform",
	context_frames: int=16,
	context_overlap: int=4,
	context_batch_size: int=1,
	interpolation_factor: int=1,
	**kwargs: Any
	) -> Pose2VideoPipelineOutput
	```

	Using a reference image (for structure) and pose images (for pose), render a video of the former in the poses of the latter, using context windowing for long-video generation.
	- The pose images here are a processed face poses. To non-processed face poses, see `vid2vid`.
	- Optionally pass `reference_pose_image` to designate the pose of `reference_image`. When not passed, the pose of `reference_image` is automatically detected.
	- Optionally pass `video_length` to use this many frames. Default is the same as the length of the pose images.