Skip to content Skip to footer

MagicDance: Practical Human Dance Video Era

Pc imaginative and prescient is without doubt one of the most mentioned fields within the AI trade, because of its potential functions throughout a variety of real-time duties. Lately, laptop imaginative and prescient frameworks have superior quickly, with fashionable fashions now able to analyzing facial options, objects, and far more in real-time eventualities. Regardless of these capabilities, human movement switch stays a formidable problem for laptop imaginative and prescient fashions. This job entails retargeting facial and physique motions from a supply picture or video to a goal picture or video. Human movement switch is extensively utilized in laptop imaginative and prescient fashions for styling photographs or movies, enhancing multimedia content material, digital human synthesis, and even producing knowledge for perception-based frameworks.

On this article, we give attention to MagicDance, a diffusion-based mannequin designed to revolutionize human movement switch. The MagicDance framework particularly goals to switch 2D human facial expressions and motions onto difficult human dance movies. Its aim is to generate novel pose sequence-driven dance movies for particular goal identities whereas sustaining the unique id. The MagicDance framework employs a two-stage coaching technique, specializing in human movement disentanglement and look components like pores and skin tone, facial expressions, and clothes. We’ll delve into the MagicDance framework, exploring its structure, performance, and efficiency in comparison with different state-of-the-art human movement switch frameworks. Let’s dive in.

As talked about earlier, human movement switch is without doubt one of the most complicated laptop imaginative and prescient duties due to the sheer complexity concerned in transferring human motions and expressions from the supply picture or video to the goal picture or video. Historically, laptop imaginative and prescient frameworks have achieved human movement switch by coaching a task-specific generative mannequin together with GAN or Generative Adversarial Networks on course datasets for facial expressions and physique poses. Though coaching and utilizing generative fashions ship passable ends in some instances, they often undergo from two main limitations. 

  1. They rely closely on a picture warping part on account of which they typically wrestle to interpolate physique components invisible within the supply picture both as a result of a change in perspective or self-occlusion. 
  2. They can not generalize to different photographs sourced externally that limits their functions particularly in real-time eventualities within the wild. 

Trendy diffusion fashions have demonstrated distinctive picture era capabilities throughout totally different situations, and diffusion fashions at the moment are able to presenting highly effective visuals on an array of downstream duties resembling video era & picture inpainting by studying from web-scale picture datasets. Owing to their capabilities, diffusion fashions is perhaps an excellent choose for human movement switch duties. Though diffusion fashions will be carried out for human movement switch, it does have some limitations both when it comes to the standard of the generated content material, or when it comes to id preservation or affected by temporal inconsistencies on account of mannequin design & coaching technique limits. Moreover, diffusion-based fashions display no important benefit over GAN frameworks when it comes to generalizability. 

To beat the hurdles confronted by diffusion and GAN based mostly frameworks on human movement switch duties, builders have launched MagicDance, a novel framework that goals to use the potential of diffusion frameworks for human movement switch demonstrating an unprecedented stage of id preservation, superior visible high quality, and area generalizability. At its core, the elemental idea of the MagicDance framework is to separate the issue into two levels : look management and movement management, two capabilities required by picture diffusion frameworks to ship correct movement switch outputs. 

The above determine offers a quick overview of the MagicDance framework, and as it may be seen, the framework employs the Steady Diffusion mannequin, and in addition deploys two further parts : Look Management Mannequin and Pose ControlNet the place the previous supplies look steerage to the SD mannequin from a reference picture by way of consideration whereas the latter supplies expression/pose steerage to the diffusion mannequin from a conditioned picture or video. The framework additionally employs a multi-stage coaching technique to study these sub-modules successfully to disentangle pose management and look. 

In abstract, the MagicDance framework is a

  1. Novel and efficient framework consisting of appearance-disentangled pose management, and look management pretraining.  
  2. The MagicDance framework is able to producing reasonable human facial expressions and human movement below the management of pose situation inputs and reference photographs or movies. 
  3. The MagicDance framework goals to generate appearance-consistent human content material by introducing a Multi-Supply Consideration Module that provides correct steerage for Steady Diffusion UNet framework. 
  4. The MagicDance framework will also be utilized as a handy extension or plug-in for the Steady Diffusion framework, and in addition ensures compatibility with present mannequin weights because it doesn’t require further fine-tuning of the parameters. 

Moreover, the MagicDance framework reveals distinctive generalization capabilities for each look and movement generalization. 

  1. Look Generalization : The MagicDance framework demonstrates superior capabilities with regards to producing numerous appearances. 
  2. Movement Generalization : The MagicDance framework additionally has the power to generate a variety of motions. 

MagicDance : Goals and Structure

For a given reference picture both of an actual human or a stylized picture, the first goal of the MagicDance framework is to generate an output picture or an output video conditioned on the enter and the pose inputs {P, F} the place P represents human pose skeleton and F represents the facial landmarks. The generated output picture or video ought to have the ability to protect the looks and id of the people concerned together with the background contents current within the reference picture whereas retaining the pose and expressions outlined by the pose inputs. 


Throughout coaching, the MagicDance framework is skilled as a body reconstruction job to reconstruct the bottom reality with the reference picture and pose enter sourced from the identical reference video. Throughout testing to realize movement switch, the pose enter and the reference picture is sourced from totally different sources. 

The general structure of the MagicDance framework will be cut up into 4 classes: Preliminary stage, Look Management pretraining, Look-disentangled Pose Management, and Movement Module. 

Preliminary Stage

Latent Diffusion Fashions or LDM symbolize uniquely designed diffusion fashions to function inside the latent area facilitated by way of an autoencoder, and the Steady Diffusion framework is a notable occasion of LDMs that employs a Vector Quantized-Variational AutoEncoder and temporal U-Internet structure. The Steady Diffusion mannequin employs a CLIP-based transformer as a textual content encoder to course of textual inputs by changing textual content inputs into embeddings. The coaching part of the Steady Diffusion framework exposes the mannequin to a textual content situation and an enter picture with the method involving the encoding of the picture to a latent illustration, and topics it to a predefined sequence of diffusion steps directed by a Gaussian methodology. The resultant sequence yields a loud latent illustration that gives a regular regular distribution with the first studying goal of the Steady Diffusion framework being denoising the noisy latent representations iteratively into latent representations. 

Look Management Pretraining

A serious problem with the unique ControlNet framework is its lack of ability to manage look amongst spatially various motions constantly though it tends to generate photographs with poses intently resembling these within the enter picture with the general look being influenced predominantly by textual inputs. Though this methodology works, it isn’t fitted to movement switch involving duties the place its not the textual inputs however the reference picture that serves as the first supply for look info. 

The Look Management Pre-training module within the MagicDance framework is designed as an auxiliary department to offer steerage for look management in a layer-by-layer method. Quite than counting on textual content inputs, the general module focuses on leveraging the looks attributes from the reference picture with the intention to reinforce the framework’s potential to generate the looks traits precisely significantly in eventualities involving complicated movement dynamics. Moreover, it is just the Look Management Mannequin that’s trainable throughout look management pre-training. 

Look-disentangled Pose Management

A naive answer to manage the pose within the output picture is to combine the pre-trained ControlNet mannequin with the pre-trained Look Management Mannequin immediately with out fine-tuning. Nevertheless, the combination may end result within the framework battling appearance-independent pose management that may result in a discrepancy between the enter poses and the generated poses. To sort out this discrepancy, the MagicDance framework fine-tunes the Pose ControlNet mannequin collectively with the pre-trained Look Management Mannequin. 

Movement Module

When working collectively, the Look-disentangled Pose ControlNet and the Look Management Mannequin can obtain correct and efficient picture to movement switch, though it’d end in temporal inconsistency. To make sure temporal consistency, the framework integrates an extra movement module into the first Steady Diffusion UNet structure. 

MagicDance : Pre-Coaching and Datasets

For pre-training, the MagicDance framework makes use of a TikTok dataset that consists of over 350 dance movies of various lengths between 10 to fifteen seconds capturing a single individual dancing with a majority of those movies containing the face, and the upper-body of the human. The MagicDance framework extracts every particular person video at 30 FPS, and runs OpenPose on every body individually to deduce the pose skeleton, hand poses, and facial landmarks. 

For pre-training, the looks management mannequin is pre-trained with a batch dimension of 64 on 8 NVIDIA A100 GPUs for 10 thousand steps with a picture dimension of 512 x 512 adopted by collectively fine-tuning the pose management and look management fashions with a batch dimension of 16 for 20 thousand steps. Throughout coaching, the MagicDance framework randomly samples two frames because the goal and reference respectively with the photographs being cropped on the similar place alongside the identical top. Throughout analysis, the mannequin crops the picture centrally as an alternative of cropping them randomly. 

MagicDance : Outcomes

The experimental outcomes carried out on the MagicDance framework are demonstrated within the following picture, and as it may be seen, the MagicDance framework outperforms present frameworks like Disco and DreamPose for human movement switch throughout all metrics. Frameworks consisting a “*” in entrance of their title makes use of the goal picture immediately because the enter, and contains extra info in comparison with the opposite frameworks. 

It’s attention-grabbing to notice that the MagicDance framework attains a Face-Cos rating of 0.426, an enchancment of 156.62% over the Disco framework, and almost 400% enhance in comparison towards the DreamPose framework. The outcomes point out the sturdy capability of the MagicDance framework to protect id info, and the seen enhance in efficiency signifies the prevalence of the MagicDance framework over present state-of-the-art strategies. 

The next figures examine the standard of human video era between the MagicDance, Disco, and TPS frameworks. As it may be noticed, the outcomes generated by the GT, Disco, and TPS frameworks undergo from inconsistent human pose id and facial expressions. 

Moreover, the next picture demonstrates the visualization of facial features and human pose switch on the TikTok dataset with the MagicDance framework having the ability to generate reasonable and vivid expressions and motions below numerous facial landmarks and pose skeleton inputs whereas precisely preserving id info from the reference enter picture. 

It’s price noting that the MagicDance framework boasts of outstanding generalization capabilities to out-of-domain reference photographs of unseen pose and kinds with spectacular look controllability even with none further fine-tuning on the goal area with the outcomes being demonstrated within the following picture. 

The next photographs display the visualization capabilities of MagicDance framework when it comes to facial features switch and zero-shot human movement. As it may be seen, the MagicDance framework generalizes to in-the-wild human motions completely. 

MagicDance : Limitations

OpenPose is a vital part of the MagicDance framework because it performs an important function for pose management, affecting the standard and temporal consistency of the generated photographs considerably. Nevertheless, the MagicDance framework nonetheless finds it a bit difficult to detect facial landmarks and pose skeletons precisely, particularly when the objects within the photographs are partially seen, or present fast motion. These points can lead to artifacts within the generated picture. 


On this article, we have now talked about MagicDance, a diffusion-based mannequin that goals to revolutionize human movement switch. The MagicDance framework tries to switch 2D human facial expressions and motions on difficult human dance movies with the particular intention of producing novel pose sequence pushed human dance movies for particular goal identities whereas conserving the id fixed. The MagicDance framework is a two-stage coaching technique for human movement disentanglement and look like pores and skin tone, facial expressions, and garments.

MagicDance is a novel method to facilitate reasonable human video era by incorporating facial and movement expression switch, and enabling constant within the wild animation era with no need any additional fine-tuning that demonstrates important development over present strategies. Moreover, the MagicDance framework demonstrates distinctive generalization capabilities over complicated movement sequences and numerous human identities, establishing the MagicDance framework because the lead runner within the discipline of AI assisted movement switch and video era. 

Leave a comment