MORPHOS

Video Results

Flip through samples — each shows the input video alongside MORPHOS's 4D mesh novel view renderings.

Why is video-to-4D generation hard?

Existing 4D generation methods perform well in narrow settings, but face three recurring limitations:

A single model across representations and long horizons

Method

Temporal Structured Latents (T-SLAT)

We extend Structured Latents (SLAT) into the temporal domain to obtain a unified 4D representation that jointly encodes geometry and appearance along time. An animated mesh sequence is normalized into a shared canonical space via a global union axis-aligned bounding box, then encoded by a sparse VAE into T-SLAT. A single T-SLAT can be decoded into meshes, 3D Gaussians, and radiance fields.

Autoregressive 4D Generation

MORPHOS factorizes 4D generation as a Markovian process: each latent z^t is conditioned on the preceding history z^<t and the current video frame. Two rectified-flow transformers generate the sparse structure (voxels) and then the T-SLAT conditioned on it. A causal attention architecture with a sliding window restricts each query to its history, enabling arbitrarily long generation and KV caching for efficient inference.

Training with Temporal-Structural Augmentation

Autoregressive models are trained on ground-truth history but must generalize to imperfect, self-generated history at inference. To bridge this gap, MORPHOS trains both flow transformers with a temporal-structural augmentation strategy.

Qualitative Results

Appearance

MORPHOS produces visually consistent, high-fidelity appearance with stable textures across the sequence, while deformation-based baselines cannot model topological changes and frame-wise TRELLIS loses temporal consistency.

Qualitative results on appearance. MORPHOS maintains stable, high-fidelity textures throughout the sequence.

Geometry

Rendered normal maps show MORPHOS generates geometrically faithful structures and handles complex topological transitions where fixed-topology methods distort or fail.

Qualitative results on geometry. MORPHOS preserves accurate shape and temporal consistency through topological changes.

Generalization to Real Domain

Quantitative Results

Evaluated against state-of-the-art video-to-4D baselines and the image-to-3D model TRELLIS. Bold = best, underline = second best per column.

Analysis

Attention captures geometric correspondence

For a query voxel (green star) in the target frame, we visualize its attention over the conditioning frames' voxel tokens. MORPHOS establishes geometric correspondences in 3D and even exploits symmetry across frames — e.g., a query on one hand attends to both hands — supporting temporally consistent voxel generation.

Minimal error accumulation over long horizons

As video length grows, baselines degrade progressively in both appearance (CLIP↑) and geometry (P2S↓). MORPHOS stays substantially more stable across the full sequence, confirming that causal generation with temporal-structural augmentation mitigates error accumulation.

Citation

@article{kwon2026morphos,
  title   = {MORPHOS: Autoregressive 4D Generation with Temporal Structured Latents},
  author  = {Kwon, Minkyung and Choi, Jinhyeok and Shin, Youngjin and
             Kim, Jaeyeong and Lee, JongMin and Kim, Seungryong},
  journal = {arXiv preprint},
  year    = {2026}
}