Animate-X: Bring Whatever(X) to Life

Animate-X: Bring Whatever(X) to Life

On the 14th, Alibaba Group and its subsidiary Ant Group unveiled Animate-X, a model capable of transforming any image into video. Similar models have emerged in the past, with some even reaching commercial service levels. What unique features does this newly published paper offer?

Why Animate-X?

Early models that used AI technology to bring life to people in photos often resulted in faces looking distorted or moving unnaturally. Some of these results were so unrealistic that they became memes (though technically using slightly different techniques). However, as technology rapidly advanced, it became possible to create videos where people move and dance naturally.
 
To create such videos, a reference image and a target pose to guide the movements are essential. Until now, reference images have focused primarily on human subjects, specifically on extracting characteristics related to human body shapes and faces. This limited the model’s versatility. The research takes a step toward generalization, aiming to animate a wide range of characters, including human-like figures and anthropomorphic animal characters. The name “Animate-X” signifies the model’s ability to animate “anything (X).”
Animate-X: Make-A-Scene model that generates images based on sketches

Make-A-Scene model that generates images based on sketches. Source: Meta AI Blog

Shall we take a look at the results? When we use a reference image of a rabbit and a guide video featuring a human, typical models generally produce a video that resembles a human figure. In contrast, the new model (Ours) generates a video where the rabbit adopts human-like poses with short limbs.
 
Why do such transformations usually occur? Let’s break it down from a technical perspective. Converting animations into a 2D pose skeleton (a simplified motion representation) has the advantage of abstracting movements that can be applied across various subjects. However, this approach often loses the finer details of the image. By using a self-driven reconstruction strategy, it becomes possible to align the reference image with the pose skeleton to represent movements more accurately. Yet, when applied to subjects with significantly different body types—such as a human and a rabbit—this method can produce unnatural results, as seen in the image above.

How Does Animate-X Work?

To overcome the above issues, the researchers designed a Pose Indicator that can guide poses from both implicit and explicit perspectives. This approach adds detailed imagery and clarifies the complex interactions between shape and pose that the traditional skeleton method lacks. Let’s look at each concept more closely:

Implicit Pose Indicator (IPI)

: Using the CLIP model, IPI extracts features from the image. CLIP is a model trained to understand the relationship between images and text, going beyond visible features to interpret the concept or meaning that an image conveys. Trained to comprehend abstract contexts by linking text descriptions with images, CLIP enables IPI to capture underlying meanings and background information not attainable through a pose skeleton alone.

Explicit Pose Indicator (EPI)

: EPI addresses mismatched situations where the reference image and target pose differ significantly. For instance, if a short-armed rabbit needs to mimic a long-armed human pose, EPI adjusts these elements to bridge the gap. By simulating these mismatches, EPI fine-tunes specific pose elements to ensure a natural alignment between character and motion. Working alongside IPI, EPI supplements the concrete movements of a pose while capturing abstract scene contexts, producing more realistic and cohesive results that are otherwise difficult to achieve with just a pose skeleton.

The Architecture of Animate-X.

The Architecture of Animate-X.

The image may appear complex, but examining each component closely reveals its structure. First, you extract information from IPI and EPI and feed it into the Latent Diffusion model (orange section in the image above) to generate the final animation. The Latent Diffusion model, an extended version of well-known generative models like Stable Diffusion, adapts specifically for character animation.
 
IPI utilizes CLIP to capture motion patterns and spatial details from the image, then merges this with pose information from DWPose* to achieve a deeper motion representation. Through Cross-Attention, these two data points are integrated to enhance the motion characteristics necessary for animation. In contrast, EPI precisely adjusts the input skeleton, converting it into a pose image and feeding it into the Latent Diffusion model to maintain motion consistency.
 
*DWPose is a lightweight model designed to quickly and accurately extract body poses, supporting natural character movements based on this pose information.
 
Animate-X leverages the collaboration of IPI and EPI to generate flexible animations that retain the pose and identity of diverse characters.
The result of using DWPose to adjust misalignment of fingers.

The result of using DWPose to adjust misalignment of fingers. Source: (Yang et al., 2024)

The results are below. Despite none of the subjects in the video being actual humans, each character adopts poses naturally. Notably, even characters without hands or legs exhibit well-adapted poses that align with the traits of the reference image. This showcases the model’s ability to generalize pose characteristics effectively.

Additionally, the researchers introduced the Animated Anthropomorphic Benchmark (A² Bench), aimed at extending evaluation standards—previously focused on human poses—to include anthropomorphic animations. Animate-X achieved a much higher quality than existing models, setting a new state-of-the-art (SOTA) standard. 

What Exactly Are IPI & EPI?

The core idea of this research lies in integrating explicit and implicit information into the generative model. In other words, it incorporates both human-understood concepts and model-learned concepts when generating images. The researchers conducted experiments to observe the results when only EPI or IPI was used. The outcomes are below. In the case without IPI, the model attempts to understand and mimic the movement, while in the case without EPI, the panda maintains its shape, albeit somewhat unnaturally.

IPI preserves the external appearance and prevents the generation of unnatural outcomes (like a plant having human hands). EPI ensures that the reference subject and pose align accurately—such as keeping a panda’s ears correctly positioned. The final result is a harmonious combination of these two elements.

Animate-X pushes the boundaries of complex character animation generation. It offers an innovative approach that maintains consistency in identity and motion across diverse character types. By combining IPI and EPI, this new model structure achieves natural and flexible movement that goes beyond simple pose estimation. This advancement sets a promising new standard, not only for animation but also for applications in gaming, virtual reality, and digital content creation.

Your AI Data Standard

LLM Evaluation Platform
About Datumo
Related Posts