MacPaw unveils GenRA, an LLM-driven framework animating rigged 3D models at ECCV
ducha_aiki · x · 2026-09-08
MacPaw Research presented GenRA at the OpenSUN3D workshop @ ECCV 2026: an end-to-end framework that uses a multimodal LLM to generate and edit skeletal animations on user-provided rigged 3D assets.
- The model is conditioned on the rig hierarchy, bind pose, multiview renders, natural-language instructions, and motion demos compactly encoded via keyframe decimation; it predicts sparse joint-transformation keyframes that are converted into smooth curves and authored back without altering geometry, skinning, or rig hierarchy
- No task-specific training or retargeting required; supports zero- and few-shot generation with localized, compositional, temporal, and quantitative refinements
- Across 1,080 blind judgments on 40 assets, GenRA achieves a 74.5% tie-aware preference rate over Puppeteer, cuts median generation time by 83% and cost by 84%
More from Multimodal
- NYC street scene made with MiniMax H3, extended via ComfyUI latent-extend node — blackdatafilms · 2026-09-08
- 8GB VRAM runs Flux and Wan 2.2 locally fine: three wrong settings, not the GPU, were the bottleneck — leonbuilds · 2026-09-08
- SuperSplat Editor 3.0 rebuilt on WebGPU, handles tens of millions of Gaussians — willeastcott · 2026-09-08
- Fixing broken text in AI slides: generate the design, strip the text, then let AI refill it — algo_diver · 2026-09-08
- VISART 2026 at ECCV calls papers as author benchmarks commercial vs local models on art catalog parsing — ducha_aiki · 2026-09-08
- This one-prompt 'cutest rescue mission' AI image is melting timelines — umesh_ai · 2026-09-08