MonoPartNeRF: Monocular Part-Aware NeRF

Updated 8 July 2026

MonoPartNeRF is a method that decomposes 3D scenes into localized neural radiance fields using monocular inputs for editable, object-centric reconstruction.
It leverages explicit part decomposition with affine transforms and per-part latent codes to achieve fine-grained control and enhanced deformation fidelity.
The framework addresses scale-depth ambiguities and improves monocular pose estimation, setting a foundation for robust NeRF-based reconstruction techniques.

Searching arXiv for MonoPartNeRF and closely related papers. Tool call: arxiv_search({"query":"MonoPartNeRF", "max_results": 10, "sort_by": "relevance"}) arXiv search results for "MonoPartNeRF":

"MonoPartNeRF: Learning Part-Aware Neural Radiance Fields from Monocular Videos" [placeholder]
... MonoPartNeRF is most naturally understood as a monocular, part-aware neural radiance field formulation that combines object-centric or canonical radiance-field reconstruction with an explicit or implicit decomposition into localized parts. In the surrounding literature, the relevant design space is defined by four recurrent ingredients: training from monocular or otherwise highly limited 2D observations; a canonical radiance field; a transformation that maps observed samples into canonical coordinates; and a part-structured representation that improves editability, controllability, or deformation fidelity. This interpretation is anchored by PartNeRF’s part-aware, image-supervised generative NeRF, PartNerFace’s monocular facial avatar reconstruction with part-based deformation, Gumbel-NeRF’s part-compositional mixture-of-experts formulation for unseen objects, and SUP-NeRF’s unification of monocular pose estimation and object-centric NeRF reconstruction (Tertikas et al., 2023, Yu et al., 15 Apr 2026, Sekikawa et al., 2024, Guo et al., 2024).

1. Problem formulation and scope

A MonoPartNeRF-style problem is defined by the conjunction of monocular ambiguity and part-structured 3D representation. PartNeRF tackles editable 3D shape generation without explicit 3D supervision and uses only posed RGB images, binary object masks, and known camera parameters during training. It explicitly excludes 3D meshes, occupancies, SDFs, part labels, or semantic segmentations. SUP-NeRF studies recovering an object’s pose, shape, and texture from a single image in real driving scenes. PartNerFace addresses animatable facial avatar reconstruction from monocular RGB videos from a single stationary camera. Gumbel-NeRF considers reconstructing and rendering novel views of unseen instances from one or a few input images (Tertikas et al., 2023, Guo et al., 2024, Yu et al., 15 Apr 2026, Sekikawa et al., 2024).

Taken together, these settings suggest that MonoPartNeRF is not defined by a single dataset protocol. A plausible implication is that the term denotes a family of object-centric or avatar-centric radiance-field methods in which monocular supervision is paired with a part decomposition to improve local control, local motion modeling, or cross-instance compositional reuse.

2. Explicit part decomposition in radiance fields

PartNeRF provides the clearest explicit part formulation. The object is represented by $M$ parts, and in all reported experiments $M=16$ . Each part is a local neural radiance field with a shape latent code $z_m^s$ , a texture latent code $z_m^t$ , a rotation $\mathbf R_m$ , a translation $t_m$ , and a scale vector $s_m$ . The transformation from global to local coordinates is

$T_m(x) = \mathbf R_m (x + t_m).$

The model uses the occupancy-based rendering form from UniSurf,

$\hat{C}(r) = \sum_{i=1}^N o_i^r \prod_{j<i} \left(1 - o_j^r \right)c_i^r,$

with

$o_i^r = 1- \exp(-\sigma_i^r \delta_i^r).$

The stated reason for adopting this occupancy parameterization is that it makes it easier to define whether a ray intersects a part and therefore easier to assign rays to parts. The central mechanism is a hard ray-to-part assignment that makes sure that the color of each ray is only determined by a single NeRF; as a result, altering one part does not affect the appearance of the others. The representation is a set of local NeRFs, one per part, each equipped with an affine transform and a spatial support ellipsoid (Tertikas et al., 2023).

Within a MonoPartNeRF interpretation, this is the canonical explicit-part design: localized coordinate frames, per-part latent variables, and rendering responsibility tied to a single part. This suggests that “part awareness” is not merely a regularizer on a monolithic field, but a structural decomposition of both geometry and appearance.

3. Canonical-space deformation and part-based motion modeling

PartNerFace shows how a part decomposition can be embedded in a canonical dynamic radiance field rather than in a static object generator. The method starts from monocular RGB video and estimated FLAME parameters per frame. It uses EMOCA to obtain per-frame shape coefficients $M=16$ 0, pose coefficients $M=16$ 1, and expression coefficients $M=16$ 2. The canonical avatar is represented as a neural radiance field defined in a canonical space. During rendering, each observed-space sample point is mapped back into canonical space using a coarse-to-fine deformation model: inverse skinning derived from FLAME, followed by a learned part-based offset field conditioned on pose and expression. The core equation is

$M=16$ 3

The canonical field itself is an MLP $M=16$ 4,

$M=16$ 5

where $M=16$ 6, $M=16$ 7, and $M=16$ 8. The method follows UNISURF and uses occupancy-based rendering with normals derived from occupancy gradients. Its part-based deformation field consists of multiple local MLPs to adaptively partition the canonical space into different parts, and the deformation of a 3D point is computed by aggregating the prediction of all local MLPs by a soft-weighting mechanism (Yu et al., 15 Apr 2026).

This establishes an important variant of MonoPartNeRF: the parts need not be separate radiance fields. They can instead be local deformation experts attached to a shared canonical radiance field. The paper’s motivation is that face motion is intrinsically part-structured because different regions have different deformation statistics, weak inter-region similarity, and strong intra-region coherence. A plausible implication is that a MonoPartNeRF architecture may place the part decomposition in the deformation model, the appearance model, or both.

4. Latent part compositionality and expert selection

Gumbel-NeRF represents a different endpoint in the design space: parts are not explicit entities with named semantics, affine transforms, or masks, but latent experts selected per 3D sample. The base conditional form is

$M=16$ 9

The model contains $z_m^s$ 0 experts $z_m^s$ 1, and the paper uses $z_m^s$ 2 experts. The decomposition is asymmetric: shape is expert-specific, texture is shared. Each expert has its own shape MLP, while the texture code and RGB head are shared across experts. All experts are evaluated first, and only then is one expert selected. The selector is density-based hindsight selection; without Gumbel noise, the final density is effectively

$z_m^s$ 3

The continuity argument is direct: if each expert density $z_m^s$ 4 is continuous in $z_m^s$ 5, then the pointwise maximum of finitely many continuous functions is also continuous. This is the basis for the claim that the method guarantees continuity in the predicted density field. By contrast, a foresight gate can produce discontinuities at routing boundaries. Qualitatively, the paper states that Coded Switch-NeRF shows “unnatural discontinuity in shape,” while Gumbel-NeRF does not. It also reports that one expert often handles the four wheels across different cars, which is presented as evidence of more consistent cross-instance decomposition (Sekikawa et al., 2024).

For MonoPartNeRF, the significance is conceptual. Part decomposition can be explicit, as in PartNeRF, or emergent, as in Gumbel-NeRF. The latter is weaker than an explicitly part-supervised or explicitly part-parameterized model, but stronger than a monolithic conditional NeRF. The literature is explicit that semantic consistency is not guaranteed: the experts are latent, reusable geometric submodels rather than guaranteed semantic parts.

5. Monocular pose estimation, canonicalization, and scale-depth ambiguity

SUP-NeRF identifies monocular pose estimation as a central systems issue for any object-centric MonoPartNeRF. Its key claim is that scale-depth ambiguity in monocular object reconstruction causes failures when the initial pose deviates moderately from the true pose. The model therefore decouples the object’s dimension estimation and pose refinement to resolve the scale-depth ambiguity, and introduces a camera-invariant projected-box representation. The geometric core is the conversion between object and camera coordinates,

$z_m^s$ 6

with

$z_m^s$ 7

The pose module uses projected 3D box corners as the pose hypothesis representation. Rotation is updated in axis-angle / Lie algebra form, while translation is updated using a relative image/depth update rather than direct metric 3D increments. The paper’s main stabilization decision is that dimension estimation is predicted once and then frozen during pose refinement. During training, the NeRF decoder uses ground-truth camera poses, not the pose estimator’s predictions, so that the NeRF decoder learns under reliable geometry. The ablation further suggests that NeRF pose optimization usually needs initial rotation error below about $z_m^s$ 8, and even then often cannot improve below about $z_m^s$ 9 (Guo et al., 2024).

This has a direct implication for MonoPartNeRF. Part-aware decomposition alone does not remove monocular degeneracies. If rays are not mapped correctly into canonical object space, localized part fields or part-specific deformation experts will still be optimized under erroneous geometry.

6. Comparative position, misconceptions, and limitations

The four neighboring lines of work delineate the practical meaning of MonoPartNeRF more clearly than any single architecture.

Work	Part mechanism	Monocular or sparse-view setting
PartNeRF	Set of local NeRFs with an affine transform and a spatial support ellipsoid	Posed RGB images, binary object masks, known camera parameters
PartNerFace	Multiple local MLPs in a part-based deformation field with soft-weighting	Monocular RGB video from a single stationary camera
Gumbel-NeRF	$z_m^t$ 0 experts with density-based hindsight selection	One-shot or few-shot input for an unseen object
SUP-NeRF	No part-aware decomposition; dedicated pose estimator plus object-centric NeRF	Single image in real driving scenes

Several misconceptions are addressed directly by this comparison. First, part-aware does not imply 3D supervision: PartNeRF is explicitly designed to work without meshes, occupancies, SDFs, part labels, or semantic segmentations (Tertikas et al., 2023). Second, part-based does not imply semantic parts: Gumbel-NeRF explicitly lets the model learn how to decompose into parts instead of giving supervision about parts, so semantic consistency is not guaranteed (Sekikawa et al., 2024). Third, monocular does not imply self-supervised: SUP-NeRF uses image crops and masks, but also uses 3D box annotations, projected 2D box corners, and ground-truth object poses during training (Guo et al., 2024). Fourth, a single global deformation network is not assumed to be sufficient for dynamic monocular avatars: PartNerFace argues that a single global MLP tends to average over heterogeneous motions across the face, blurring subtle movements and fine geometry (Yu et al., 15 Apr 2026).

The limitations also differ systematically. Gumbel-NeRF evaluates all experts before selection and is therefore slower than a foresight-gated MoE. Its decomposition is more geometric than fully appearance-compositional because texture is shared while only shape is expert-specific (Sekikawa et al., 2024). SUP-NeRF is more supervised on the pose side than many monocular object-NeRF approaches (Guo et al., 2024). PartNerFace is specialized to facial capture and FLAME-guided deformation (Yu et al., 15 Apr 2026). PartNeRF is optimized for editable shape synthesis rather than monocular pose estimation or dynamic reconstruction (Tertikas et al., 2023).

In aggregate, MonoPartNeRF is best interpreted as the intersection of these concerns: monocular or severely view-limited supervision, canonical object or avatar space, explicit or latent part decomposition, and a rendering or deformation pipeline that preserves local responsibility. This suggests a research program rather than a single fixed recipe: explicit local radiance fields for editability, local deformation experts for motion fidelity, expert competition for latent compositionality, and dedicated pose canonicalization to manage monocular ambiguity.