From Volumetric Fields to Explicit Meshes in 3D Generation

31 concepts23 papers1 proposal

When generating or reconstructing 3D assets from 2D images or text prompts, choosing how to represent the 3D geometry is a fundamental challenge. Neural implicit representations, such as Neural Radiance Fields (NeRFs) and continuous Signed Distance Functions (SDFs), offer unparalleled optimization flexibility. Because they represent geometry as continuous fields, gradient descent can easily alter topology—such as carving out holes or extending limbs—without getting stuck in catastrophic local minima. However, this continuity comes at a cost: implicit fields are computationally expensive to render, inherently struggle to represent high-frequency textures, and often produce blurry surfaces.

Conversely, explicit representations like triangular meshes, tetrahedral grids, and point clouds are the gold standard for traditional graphics pipelines. They allow for highly efficient rasterization and can lock down razor-sharp geometric boundaries and textures. Yet, optimizing an explicit mesh from scratch is notoriously difficult; if the initial topology is wrong (e.g., a sphere trying to become a donut), the optimization will fail because tearing or merging discrete faces is non-differentiable.

To resolve this dilemma, a powerful paradigm has emerged: a two-stage optimization pipeline that transitions from an implicit volumetric representation to an explicit surface representation. By splitting the generation process into two distinct phases, models exploit the unique advantages of each representation exactly when they are needed.

The Coarse-to-Fine Transition Pipeline

The core intuition behind the implicit-to-explicit transition is a coarse-to-fine progression. In the first stage, the system initializes a volumetric field to discover the global structure and topology of the object. Once the coarse geometry is established, the pipeline transitions to an explicit mesh or grid for high-resolution refinement.

For instance, Guocheng Qian et al. demonstrate in Magic123 that this shift enables the generation of high-resolution textures and intricate geometric details while remaining memory-efficient. Similarly, Jianfeng Zhang et al. (AvatarStudio) and Xiao Han et al. (HeadSculpt) apply this strategy to 3D avatar creation, showing that while volumetric fields are effective for initial geometry discovery, explicit meshes are essential for capturing sharp textures and fine-grained details.

This transition also acts as a regularizer against optimization artifacts. Dengsheng Chen et al. (Real3D) observe that shifting from implicit to explicit rendering during training significantly improves geometric sharpness and consistency, preventing "neural scene degeneration" that often results in toy-like, unrealistic outputs.

[Stage 1: Implicit Field] ──(Differentiable Bridge)──> [Stage 2: Explicit Mesh]
  - Continuous Optimization                              - Discrete Rasterization
  - Fluid Topology Discovery                             - Sharp High-Frequency Textures
  - Smooth Gradient Descent                              - Downstream Engine Compatibility

Differentiable Bridges via Tetrahedral Grids

The core technical challenge in this paradigm is making the transition from a continuous implicit field to a discrete explicit structure differentiable. If the extraction step is a hard threshold (like standard Marching Cubes), gradients from the final explicit rendering stage cannot flow back to update the underlying geometry.

To bridge this gap, researchers frequently turn to Differentiable Marching Tetrahedra (DMTet). By defining a Signed Distance Function over a deformable tetrahedral grid, the extraction of the triangular mesh becomes differentiable with respect to both the SDF values and the grid vertex positions.

Shangzhe Wu et al. (MagicPony) utilize this hybrid approach to optimize for complex topological changes during training while retaining the rendering efficiency of traditional meshes. This mechanism is crucial for downstream compatibility; as Ye Chen et al. (MeshAvatar) note, extracting explicit triangular meshes via DMTet allows learned digital characters to be directly integrated into traditional graphics engines and game development workflows.

During the refinement stage, the tetrahedral grid itself can be progressively scaled. Baorui Ma et al. (GeoDream) and Jinbo Wu et al. (HD-Fusion) use a neural SDF for coarse shape acquisition before transitioning to a deformable marching tetrahedra representation, allowing for finer control over surface geometry and high-frequency textures.

Similarly, Xuanyu Yi et al. (Diffusion Time-step Curriculum) evolve their representation from a low-resolution volumetric field with masked hash-encodings to a high-resolution tetrahedral grid, ensuring robust global geometry before mapping fine textures. This progression is also highly effective for capturing complex, non-rigid structures; Yuan Zhong et al. (MoReMouse) leverage a volumetric-to-tetrahedral mesh transition to accurately reconstruct thin extremities and fine-scale topological details of laboratory mice.

Alternative Explicit Targets: Points and Gaussians

While triangular meshes are the most common destination, the implicit-to-explicit transition strategy extends to other discrete representations, such as point clouds and 3D Gaussians, depending on the desired rendering properties.

For example, Junshu Tang et al. (Make-It-3D) transition from volumetric NeRFs to discrete point-cloud representations. They show that this shift recovers sharp textures that are typically blurred out by continuous volumetric optimization, proving that hybrid representation pipelines are essential for high-frequency details.

Alternatively, the authors of DressRecon transition from an implicit signed distance field to an explicit 3D Gaussian representation. This hybrid approach leverages the strengths of implicit models to establish a watertight global topology and the strengths of explicit point-based models to achieve photorealistic, high-frequency surface textures during final refinement.

Why Disentangling the Optimization Works

The fundamental reason why the implicit-to-explicit transition works so well is that it disentangles the optimization objectives. Trying to optimize global topology and high-frequency surface details simultaneously creates severe gradient conflicts.

By separating these tasks, pipelines can first solve the macro-level geometry and then lock it down. As Jimin Xu et al. (MPOD123) and Jingxiang Sun et al. (DreamCraft3D) explain, transitioning from an implicit neural surface to an explicit mesh-based representation allows a progression from coarse topology discovery to fine-grained geometric refinement, which effectively mitigates multi-faced artifacts like the notorious Janus problem. This staging ensures that high-frequency textures are injected precisely within a structurally sound outline. Xiangjun Gao et al. (ConTex-Human) reinforce this intuition, using a continuous field to establish basic volume and pose before refining surface details and normal maps via an explicit mesh.

This decoupling of smooth geometry from sharp details is mathematically powerful across domains. In surface reconstruction, Jiaqi Leng 0002 et al. (FNIN) employ a two-stage hybrid approach that decouples smooth surface geometry from sharp discontinuities to ensure mathematical consistency. Even in physics-based multi-modal tasks, Xutong Jin et al. (DiffSound) demonstrate that integrating neural implicit representations with differentiable mesh extraction allows backpropagating errors from simulated physical feedback (such as acoustic signals) to precisely refine an object's final shape.

Key Foundational Work

AvatarStudio: High-fidelity and Animatable 3D Avatar Creation from Text (2023) — contributed 2 concept elements to this direction.
Physics‐Based Inverse Rendering using Combined Implicit and Explicit Geometries (2022) — contributed 2 concept elements to this direction.
HeadSculpt: Crafting 3D Head Avatars with Text (2023) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_implicit_to_explicit_representation_tran_4de5bf91,
  title = {implicit-to-explicit representation transition},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/implicit-to-explicit-representation-transition/rdr_4de5bf914fc5459eb6fb6126514785e9},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_implicit_to_explicit_representation_tran_4de5bf91_v1,
  title = {implicit-to-explicit representation transition},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/implicit-to-explicit-representation-transition/rdr_4de5bf914fc5459eb6fb6126514785e9/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake