blendi-remade/prompt2rig

prompt2rig is worth a look if you already work with Mixamo rigs and want a fast way to sketch motion from a sentence. Its main catch is licensing: the high-quality path is explicitly research-only.

The problem

Animators and modders who need a specific motion usually start from a description in their head, then face a choice: hand-key the pose sequence, search stock libraries, or record reference video. Each step separates the idea from a rigged result, often by hours. prompt2rig tries to collapse that into one loop: type the motion, get a generated video, recover 3D body and finger motion from it, and see it on a Mixamo character in the browser. You can then export a .glb for Blender or Unity.

What's different

The pipeline splits into a fast preview and a slower quality pass. A hosted video model on fal generates the clip, and MediaPipe runs the live preview in the browser immediately. Meanwhile, a local GPU worker runs GVHMR for body motion and WiLoR for hands, then retargets the result onto the rig. The README reports roughly 8–9s for a 5s FLUX 3 Max clip and ~3.5–4s for the HQ pass on an RTX 4090.

The practical benefit is that you get something to look at while the accurate pass finishes, rather than waiting on one slow run. The trade-off is that the useful path depends on your own NVIDIA GPU and several large checkpoint downloads; the live preview alone is less accurate. The README also warns that hands without MANO fall back to MediaPipe, which has rougher depth. The biggest constraint is license mixing: GVHMR, SMPL-X, WiLoR and MANO are non-commercial, so the HQ path is research-only even though the project's own code is Apache-2.0. A commercial product would need to replace or license those pieces.

Who should try it

Good fit: a game modder or indie animator on Windows with an NVIDIA GPU who wants to prototype motion phrases and bring a Mixamo character to life quickly. The project ships three built-in Mixamo rigs and accepts any T-pose .fbx or .glb from mixamo.com, so it fits an existing rig workflow.

Poor fit: anyone needing a commercially shippable pipeline without swapping components. The setup instructions require a fal API key for generation, and the HQ path needs Windows, an NVIDIA GPU, uv, ffmpeg and registration-gated SMPL-X and MANO downloads. A viewer who just wants a quick live preview can run the web app alone, but that path is not the accurate one.

Requirements and limitations

Generation depends on a fal API key; the server proxies it so the key stays server-side, but you are using a paid third-party service. The HQ pass requires Windows and an NVIDIA GPU plus roughly 5 GB of checkpoints, and it is tested only on Windows 11 with an RTX 4090. SMPL-X and MANO require separate registration and manual file placement, which adds setup friction before any motion result. Ongoing upkeep includes keeping the worker warm and managing model licenses; the exported .glb may carry the same non-commercial restrictions if it embeds data derived from those models.

HotRepo's take

HotRepo recorded 399 total stars on the NEW-REPOSITORY board for JavaScript on 2026-10-25, two days after its first tracking date. That is a snapshot, not a daily growth figure, and the comparable history is insufficient to judge persistence. Before investing time, check whether your intended use fits the non-commercial GVHMR/SMPL-X/WiLoR/MANO terms.