FollowYourPose_v1

Model Card for FollowYourPose V1

<!-- Provide a quick summary of what the model is/does. -->

Model Details

Model Description

<!-- Provide a longer summary of what this model is. -->

The authors note in the assoicated paper:

Pose-Guided Text-to-Video Generation. We propose an efficient training scheme to empower the ability of the pretrained text-to-image model (i.e., Stable Diffusion) to generate pose-controllable character videos with minimal data requirements. We can generate various high-definition pose-controllable character videos that are well-aligned with the pose sequences and the semantics of text prompts.

• Developed by: Yue Ma, Yingqing He, Xintao Wang, Ying Shan, Xiu Li, Qifeng Chen
• Shared by: Yue (Jack) Ma
• Model type: Text-to-Video
• Language(s) (NLP): [More Information Needed]
• License: MIT
• Finetuned from model: Stable Diffusion T2I

Model Sources

<!-- Provide the basic links for the model. -->

• Repository: GitHub Repo
• Paper: Associated Paper
• Demo: HF Demo

Uses

<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->

Direct Use

<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->

Text to Video

Downstream Use

• Character replacement
• Background change
• Style transfer
• Multiple characters

Out-of-Scope Use

<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->

Note: This section as adpated from the Stable Diffusion v1 model card. The model was not trained to be factual or true representations of people or events, and therefore using the model to generate such content is out-of-scope for the abilities of this model.

Bias, Risks, and Limitations

<!-- This section is meant to convey both technical and sociotechnical limitations. -->

Note: This section is adpted from the DALLE-MINI model card.

The model should not be used to intentionally create or disseminate videos that create hostile or alienating environments for people. This includes generating videos that people would foreseeably find disturbing, distressing, or offensive; or content that propagates historical or current stereotypes.

Recommendations

<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->

Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.

How to Get Started with the Model

Use the code below to get started with the model.

from diffusers import DiffusionPipeline

pipeline = DiffusionPipeline.from_pretrained("YueMafighting/FollowYourPose_v1")

Training Details

Training Data

<!-- This should link to a Data Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->

Stage 1 Training Data: Pose-Controllable Text-to-Image Gen- eration. The authors note in the assoicated paper that they colected

the human skeleton images in the LAION by Mpose, only retaining images that could be detected more than 50% of the key point which formed a dataset named LAION-Pose from.

The autheors note in the associated paper about LAION-Pose:

This dataset contains diverse human-like characters with various back- ground contexts.

Stage 2 Training Data: Video Generation via Pose-free Videos.

However, the stage 1 model can generate similar pose videos yet the background is inconsistent. Thus, we further finetune the model from our first stage on the pose- free video dataset HDVLIA This dataset contains con- tinuous in-the-wild video text pairs.

Training Procedure

<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->

Preprocessing

The authors note in the assoicated paper:

We learn a zero-initialized convolutional encoder to encode the pose information We finetune the motion of the above network via a pose-free video dataset by adding the learnable temporal self-attention and reformed cross-frame self-attention blocks.

Training Hyperparameters

• Training regime: [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->

Speeds, Sizes, Times [optional]

<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->

[More Information Needed]

Evaluation

<!-- This section describes the evaluation protocols and provides the results. -->

Testing Data, Factors & Metrics

Testing Data

<!-- This should link to a Data Card if possible. -->

[More Information Needed]

Factors

<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->

[More Information Needed]

Metrics

<!-- These are the evaluation metrics being used, ideally with a description of why. -->

The authors note in the assoicated paper:

• CLIP score (CS): CLIP score or video-text alignment. We compute CLIP score for each frame and then average them across all frames. The final CLIP score is calculated on 1024 video samples.
• Quality (QU): We conduct the human evaluation of videos’ quality across a test set containing 32 videos. In detail, we display three videos in random order and request the evaluators to iden- tify the one with superior quality
• Pose Accuracy (FQ): We regard the input pose sentence as ground truth video and evaluate the average precision of the skeleton on 1024 video samples. For a fair comparison, we adopt the same pose detector for both the processing of LAION-Pose collecting and evaluation.
• Frame Consistency (FC): Following we report frame consistency measured via CLIP cosine similarity of consecutive frames

Results

Summary

Model Examination [optional]

<!-- Relevant interpretability work for the model goes here -->

[More Information Needed]

Environmental Impact

<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->

Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).

• Hardware Type: 8 NVIDIA Tesla 40G-A100 GPUs
• Hours used: 48
• Cloud Provider: [More Information Needed]
• Compute Region: [More Information Needed]
• Carbon Emitted: [More Information Needed]

Technical Specifications [optional]

Model Architecture and Objective

The authors note in the assoicated paper:

To make the pre-trained T2I model suitable for video inputs, we make several key modifications. Firstly, we add extra temporal self-attention layers, to the stable diffusion network. Secondly, inspired by the recent one-shot video generation model, we reshape the attention to cross-frame attention for better content consistency.

Compute Infrastructure

[More Information Needed]

Hardware

[More Information Needed]

Software

[More Information Needed]

Citation [optional]

<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->

BibTeX:

@article{ma2023follow,
  title={Follow Your Pose: Pose-Guided Text-to-Video Generation using Pose-Free Videos},
  author={Ma, Yue and He, Yingqing and Cun, Xiaodong and Wang, Xintao and Shan, Ying and Li, Xiu and Chen, Qifeng},
  journal={arXiv preprint arXiv:2304.01186},
  year={2023}
}

APA:

[More Information Needed]

Glossary [optional]

<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->

[More Information Needed]

More Information [optional]

[More Information Needed]

Model Card Authors [optional]

Yue Ma in collaberation with Ezi Ozoani and the Hugging Face team.

Model Card Contact

If you have any questions or ideas to discuss, feel free to contact Yue Ma or Yingqing He or Xiaodong Cun.