Alibaba's LHM turns single photos into animatable 3D avatars in seconds

Turning a single 2D photo into a fully rigged, animatable 3D avatar usually takes one of two painful paths: book a trip to an expensive multi-camera photogrammetry rig, or wait hours for an optimization algorithm to crunch through a video clip.

Researchers at Alibaba just introduced LHM, short for Large Animatable Human Reconstruction Model. The system can take a single still image of a person and spit out a fully animatable 3D digital human in just a few seconds.

Ditching slow optimization for raw speed

Reconstructing an animatable 3D human from just one picture is notoriously hard because the model has to magically decouple geometric shape, surface appearance, and movement deformation with zero depth cues.

Most existing methods either rely on static 3D scans that fail to generalize to real-world photos, or depend on computationally brutal video optimization pipelines. LHM takes a completely different route by handling the whole process in a single feed-forward pass.

Under the hood, the system represents the human body using 3D Gaussian splatting. It uses a Multimodal Body-Head Transformer (MBHT) to extract tokens from the input picture and fuse geometric body positional features with appearance details via attention mechanisms, instantly decoding them into Gaussian parameters.

No post-processing for faces or hands

Face distortion and blurry hands have long been the bane of automated 3D avatar generation, often forcing creators to patch facial meshes by hand after the initial generation.

To keep identity intact, LHM introduces a head feature pyramid encoding scheme that aggregates multi-scale features around the facial area. The team notes that the model generates realistic avatars in seconds without requiring any manual clean-up or post-processing for the face and hands.

That said, the results are not entirely flawless yet. The authors note that test animations can show white occlusion artifacts along the edges due to mask cropping, and certain motion estimations still suffer from minor jittering.

Why it matters

We think the real breakthrough here is shrinking the turnaround time for functional 3D avatars from hours to mere seconds.

For game developers, virtual creators, and digital commerce platforms, being able to turn a user selfie into a drivable 3D asset on the fly eliminates massive technical overhead. Instead of building complex rigging pipelines, studios could let users drop their real-world likeness into interactive worlds almost instantaneously.

Across the broader AI landscape, this signals a shift from static 3D asset generators toward truly dynamic, animatable representations. By proving that a transformer-based feed-forward network can handle complex body deformations via 3D Gaussian splatting, Alibaba is putting pressure on slower optimization-based workflows.

Still, the visible mask artifacts and motion jitter show that inferring unseen geometry from a single shot remains tricky. It looks promising enough for real-time consumer apps, but high-end production pipelines will likely demand cleaner edge handling and smoother tracking before adopting it fully.