MorphGS: Morphology-Adaptive Articulated 3D Motion Transfer from Videos

How to read this

Category

video-to-3D motion retargeting, analysis-by-synthesis framework

Contributions

• Recasts retargeting as target-driven analysis-by-synthesis: optimizes the target character's own morphology and pose directly from image-space supervision, instead of reconstructing the source in 3D first
• Introduces a rig-coupled morphology parameterization that factors time-invariant character identity (bone lengths, scale, rest-pose offsets) from time-varying per-frame pose so shape and motion do not become entangled during optimization
• Combines dense 2D-3D semantic correspondences with diffusion-based novel-view synthesis to give the optimizer both part-level anchors and multi-view guidance under monocular ambiguity
• Validated on synthetic benchmarks and real-world video, showing improvements over reconstruct-then-retarget baselines without needing category-specific parametric templates

Context

Standard video-to-rig retargeting pipelines reconstruct an intermediate 3D representation from the source video, then transfer the recovered motion onto the target, which propagates reconstruction error and ties the method to parametric templates for well studied categories like humans and quadrupeds. MorphGS instead optimizes the target character's own rig-coupled morphology and pose directly against the source video through differentiable rendering, avoiding that reconstruction step. This entry lists no builds_on ancestors in the archive, but it explicitly positions itself against reconstruct-then-retarget methods and prior analysis-by-synthesis work in the retargeting literature.

Correctness

The claims rest on synthetic benchmark comparisons plus qualitative real-world video results against baseline retargeting methods, described as consistent improvements. The abstract and introduction do not specify which baselines or metrics, so how large the improvement is, and how it holds up on categories with extreme morphology mismatch such as quadruped to biped, is not established from what was read. It is also inherently monocular, so it inherits the depth and occlusion ambiguity any single-camera method faces despite the added multi-view synthesis signal.

Clarity

The introduction and abstract are clearly written and the core idea, optimize the target instead of reconstructing the source, is easy to grasp even without a background in differentiable rendering. A rigger will need to read the method section for how the rig-coupled parameterization concretely maps to bone lengths and rest-pose offsets, since that detail sits past where this reading pass stopped.

How to read it

First pass, read the abstract and Figure 1 to confirm this is about direct target-mesh optimization rather than motion capture transfer, and skim the contribution bullets. Second pass, read the method section on the rig-coupled morphology parameterization and how the 2D-3D correspondence and novel-view synthesis losses combine, since that is the actual technical contribution. Third pass, if implementing, check the experiments section for exactly which categories and rigs were tested and whether optimization runs per-video or is amortized, since that determines whether this is production-viable or a research prototype.