Large-scale Codec Avatars: The Unreasonable Effectiveness of Large-scale Avatar Pretraining

Dimensions overall score 7.0

• Large-Scale Codec Avatars (LCA) uses a pre/post-training paradigm for 3D avatar modeling at scale, pretraining on 1M in-the-wild videos then post-training on high-quality curated data.method 95%
  Evidencepartial

  we present, for the first time, a pre/post-training paradigm for 3D avatar modeling at scale: we pretrain on 1M in-the-wild videos to learn broad priors over appearance and geometry, then post-train on high-quality curated data to enhance expressivity and fidelity.

  Implicationpartial

  Explicitly stated in the abstract with specific numbers and methodology description

  Verificationpartial

  partial

• LCA generalizes across hair styles, clothing, and demographics while providing precise facial expressions and finger-level articulation control.result 90%
  Evidencepartial

  LCA generalizes across hair styles, clothing, and demographics while providing precise, fine-grained facial expressions and finger-level articulation control, with strong identity preservation.

  Implicationpartial

  Directly stated in abstract as a key capability of the method

  Verificationpartial

  partial

• LCA shows emergent generalization to relightability and loose garment support to unconstrained inputs without direct supervision.result 85%
  Evidencepartial

  Notably, we observe emergent generalization to relightability and loose garment support to unconstrained inputs

  Implicationpartial

  Explicitly stated as an observed emergent property, though 'emergent' suggests it wasn't directly trained for

  Verificationpartial

  partial

• LCA demonstrates zero-shot robustness to stylized imagery despite the absence of direct supervision.result 85%
  Evidencepartial

  and zero-shot robustness to stylized imagery, despite the absence of direct supervision.

  Implicationpartial

  Directly stated as a capability, though 'zero-shot' suggests it wasn't explicitly trained for this

  Verificationpartial

  partial

• Multi-view studio data enables high-fidelity modeling but struggles to generalize to real-world data due to limited scale and domain gap.limitation 90%
  Evidencepartial

  multi-view studio data enables high-fidelity modeling of humans with precise control over expressions and poses, but it struggles to generalize to real-world data due to limited scale and the domain gap between the studio environment and the real world.

  Implicationpartial

  Directly stated as a limitation of existing approaches in the abstract

  Verificationpartial

  partial

• Large-scale avatar models trained on millions of in-the-wild samples often produce low-quality avatars due to inherent 3D ambiguities.limitation 90%
  Evidencepartial

  recent large-scale avatar models trained on millions of in-the-wild samples show promise for generalization across a wide range of identities, yet the resulting avatars are often of low-quality due to inherent 3D ambiguities.

  Implicationpartial

  Directly stated as a limitation of existing approaches in the abstract

  Verificationpartial

  partial

• LCA enables efficient inference through feedforward processing for world-scale populations.technical 85%
  Evidencepartial

  a high-fidelity, full-body 3D avatar model that generalizes to world-scale populations in a feedforward manner, enabling efficient inference.

  Implicationpartial

  Directly stated in abstract, though 'efficient' is somewhat qualitative

  Verificationpartial

  partial

• The approach is inspired by the success of large language models and vision foundation models.method 90%
  Evidencepartial

  Inspired by the success of large language models and vision foundation models

  Implicationpartial

  Directly stated in abstract as inspiration for the methodology

  Verificationpartial

  partial

Competitive landscape

A large-scale pre-trained 3D avatar model that achieves high fidelity and generalization to real-world data, enabling efficient inference for diverse applications.