Möbius: A unified AI model that generates, understands, and edits 3D motion

Möbius is a unified motion-language framework that handles 3D human motion generation, captioning, completion, and editing all within a single architecture. Instead of juggling fragmented toolchains, you can generate 3D character animations from a prompt, turn existing raw motion back into text descriptions, or prompt the model to patch and tweak animations on the fly.

Beyond standard text-to-motion, Möbius naturally handles incomplete motion sequences. You can feed it an unfinished clip and prompt it on how to fill in the missing frames, or use plain instructions to edit specific actions without re-rigging from scratch.

How it works under the hood

Existing unified motion models are rarely truly unified. Most offload understanding to an LLM while leaving motion synthesis to a completely detached latent diffusion module.

Möbius changes this by dragging the entire diffusion process into the exact same multimodal space used for motion comprehension. Clean motion sequences, noisy motion inputs, and text prompts all participate together in a shared attention mechanism, using role-specific parameters to route generation and understanding streams.

Crucially, Möbius works directly on raw, continuous frame-level motion. It throws out discrete tokenization and removes the need for a pretrained motion VAE entirely, keeping the representation clean and expressive.

Benchmarks and current availability

The researchers scaled Möbius across two configurations: a lightweight 0.6B model and a larger 1.7B variant trained on roughly 1,700 hours of motion data.

On the standard HumanML3D benchmark, the 0.6B model scored a top-1 retrieval rate of 0.8209 and drove the FID score down to 6.1291, outperforming previous unified models like MotionGPT3. The 1.7B model demonstrated that continuous multi-task motion scaling holds up nicely across diverse movement datasets.

The paper is currently live on arXiv along with a demo showcase page. Note that the project's GitHub repository states the code is undergoing internal review, meaning source files and model weights are not immediately downloadable just yet.

Why it matters

For 3D animators, game developers, and teams using video models for previs, Möbius hits a major workflow bottleneck. Motion synthesis has always been tricky because generation and conversational editing were separate beasts. Having one network understand what a character is doing while simultaneously tweaking its movement via instructions makes animation iterations significantly cheaper.

We think dropping the motion VAE in favor of continuous frame-level diffusion inside a language model architecture is an exceptionally clean architectural bet. The only catch right now is availability. Keep this on your radar, track the project page for the code release, and see how easy it is to integrate into your DCC software once the repo opens up.