HuMo | high fidelity controllable human motion

HuMo is a human centric video generator that finally treats motion, identity, and timing with the care filmmakers expect. You can guide a performance with text, an image reference, audio, or any mix of the three, and the system keeps the subject consistent while staying loyal to the instructions. In practice that means dance, acting beats, or dialogue synced gestures read as a single performance rather than a collage of almost right frames. It is not trying to wing the result from a short prompt. It anchors on clear inputs and then builds a sequence that respects them. For teams who need to see an idea before they spend on sets and time, that shift matters. You can put a look and a body in motion, match rough timing to a line, and decide whether a choice plays on screen without asking the audience to forgive the physics.

For production the model lands in a useful spot between speed and quality. There are two public checkpoints that map well to real work. A lighter 1.7B model that can create 480p sequences on a single 32 GB GPU for fast exploration, and a stronger 17B model that targets 720p with better fidelity when you are aiming for finals and hero shots in a pitch reel. The project includes a straightforward quick start, support for text plus image guidance when you have a face or costume to preserve, and text plus audio when you want motion that tracks a voice. The scheduler and configuration options are documented with sane defaults, so you can push frames without babysitting every parameter. As always, quality follows inputs. Give it a clean reference, a concrete prompt, and audio that matches the intent, and you will get motion that feels directed instead of random.

Licensing is practical for many studios. The Hugging Face model card lists Apache 2.0, and the repository provides runnable code and scripts. That is permissive enough for most experiments and a wide range of commercial work, provided you still do the basics that protect your project and your performers. Clear likeness and voice rights when you use real people. Keep a simple log of prompts, seeds, and versions so you can reproduce a shot or answer questions later. Test dialogue heavy scenes early, since face and hand fidelity matter most when the camera lingers. When you move from exploration to delivery, route the pass through editorial and sound the way you normally would. HuMo is a fast way to reach a believable motion prototype, and the best results come when you pair that speed with the same judgment you use on set.

Sources

• Project page: https://phantom-video.github.io/HuMo/
• Hugging Face model card and weights: https://huggingface.co/bytedance-research/HuMo
• GitHub repository: https://github.com/Phantom-video/HuMo
• Paper: https://arxiv.org/abs/2509.08519