HumanNOVA: Single-Image 3D Avatars

• HumanNOVA is a polysemous term with varied meanings, primarily referring to a feed-forward model for reconstructing photorealistic 3D human avatars from a single image.
• The approach fuses image tokens from DINOv2 and mesh tokens from PTv3 with a triplane representation, achieving high fidelity in geometry and texture through volume rendering.
• It employs a scalable data pipeline, generating over 2.6M training images from synthetic and real assets to enable rapid (<1 second) inference on modern GPUs.

HumanNOVA is a polysemous research label rather than a single uniformly accepted term. In the most explicit titled usage, it denotes a feed-forward large reconstruction model for photorealistic 3D human avatar modeling from a single RGB image; in adjacent literature, related usages designate an umbrella for structure-aware human novel view synthesis, a colloquial reference to the NOVA-Human anime reconstruction dataset, a synthesis of human de novo gene emergence, and a NOVA-based blueprint for human-amplified knowledge discovery (Hu et al., 1 Jun 2026, Dey et al., 2024, Wang et al., 2024, Ruiz-Orera et al., 2015, Avestimehr et al., 12 May 2026).

1. Terminology and referential scope

The term has materially different meanings across subfields. Among the sources considered here, only one paper uses HumanNOVA as the formal title of a method; the others use the label as a broad thematic shorthand, a dataset reference, or a derived conceptual framing.

Usage	Meaning	Source
HumanNOVA	Feed-forward large reconstruction model for single-image 3D human avatar modeling	(Hu et al., 1 Jun 2026)
HumanNOVA	Broadly understood as human-centric novel view synthesis with explicit, structure-aware human modeling	(Dey et al., 2024)
HumanNOVA / NOVA-Human	Refers to the NOVA-Human dataset for full-body anime reconstruction from non-overlapped views	(Wang et al., 2024)
HumanNOVA	Human de novo gene emergence and characterization	(Ruiz-Orera et al., 2015)
HumanNOVA	NOVA-based blueprint for human-amplified knowledge discovery	(Avestimehr et al., 12 May 2026)

In computer graphics and 3D vision, the most fully specified use is the 2026 single-image avatar model. That usage is therefore the central referent in what follows, with the other meanings treated as related but distinct terminological branches.

2. HumanNOVA as a single-image 3D human avatar model

HumanNOVA is defined as a feed-forward large reconstruction model specialized for single-image 3D human avatar modeling. The task is to reconstruct a static, photorealistic 3D avatar from a single RGB image and render it from any viewpoint. The paper frames the problem as ill-posed because most geometry and appearance are occluded in single-view input, while existing human-specific datasets are too small to train high-capacity feed-forward LRMs. HumanNOVA addresses this by combining a scaled data pipeline with a model conditioned jointly on image tokens and a simplified SMPL prior (Hu et al., 1 Jun 2026).

The data generation pipeline follows two complementary strategies. The first uses high-quality, rigged human assets from SynBody and animates them with daily-life poses sampled from AMASS. The second uses multi-camera datasets, specifically DNA-Rendering and MVHumanNet, and fits them with 3D Gaussian Splatting initialized from SMPL-X vertices. These strategies produce a total of 100k assets, comprising ~78k synthetic assets generated and ~22k real-world assets fitted and re-rendered, with ~2.6M training images. The synthetic branch renders on average 26 views per asset, with camera positions uniformly sampled on a sphere, azimuth 0°–360°, elevation −45°–60°, and no frontal-view bias in training.

The SMPL prior is not the output representation; it is a conditioning prior. At test time, the input consists of a single RGB image $I \in \mathbb{R}^{H \times W \times 3}$ and a simplified human mesh estimated by Humans in 4D. The paper states that this coarse SMPL prior provides body pose and shape without detailed garments or appearance, and that detailed clothing, hair, and accessories are represented by learned appearance in the triplane/NeRF. Ablations show that including mesh tokens improves LPIPS by approximately 2.3% relative versus removing the mesh prior.

A central claim of the method is the joint attainment of photorealism, universality, and speed. “Photorealistic” refers to high-fidelity geometry and texture under volume rendering and multi-view supervision. “Universal” refers to generalization across identities, body shapes, clothing, and viewpoints, supported by the mixed synthetic-real training distribution. “Rapid” refers to less than one second inference on modern datacenter GPUs and the absence of any test-time optimization.

3. Architecture, rendering, and optimization

HumanNOVA uses a token-conditioned multi-modal architecture. The image encoder is DINOv2, which produces visual tokens $f_i \in \mathbb{R}^{N_i \times d}$. The mesh encoder is PTv3 (Point Transformer v3), which encodes the estimated SMPL mesh into mesh tokens $f_m \in \mathbb{R}^{N_m \times d}$. These tokens are fused with learnable triplane tokens $T \in \mathbb{R}^{3hw \times d}$ through cross-attention blocks (Hu et al., 1 Jun 2026).

The fusion sequence is specified as

$$L_l = \mathrm{CrossAttn}(q = f_i \Vert f_m, kv = T_l),$$

$$L_l = \mathrm{CrossAttn}(q = L_l, kv = f_i \Vert f_m),$$

$T_{l+1} = \mathrm{CrossAttn}(q = T_l, kv = L_l),$

with attention defined by

$$\mathrm{Attn}(Q, K, V) = \mathrm{softmax}(QK^\top/\sqrt{d})V.$$

The 3D representation is a tri-plane with three orthogonal feature planes $F_{xy}$, $F_{yz}$, and $f_i \in \mathbb{R}^{N_i \times d}$0, each of resolution $f_i \in \mathbb{R}^{N_i \times d}$1. For a 3D point $f_i \in \mathbb{R}^{N_i \times d}$2, features are sampled by bilinear interpolation as

$f_i \in \mathbb{R}^{N_i \times d}$3

A NeRF MLP $f_i \in \mathbb{R}^{N_i \times d}$4 predicts color and density:

$f_i \in \mathbb{R}^{N_i \times d}$5

The MLP has 10 layers, width 60, SiLU activation, and 128 samples per ray. Rendering follows standard volume rendering: $f_i \in \mathbb{R}^{N_i \times d}$6` When required for geometry evaluation, an isosurface is extracted from the density field by marching cubes.

Training is end-to-end with multi-view supervision and patch-level rendering. The full objective is

$f_i \in \mathbb{R}^{N_i \times d}$7

with $f_i \in \mathbb{R}^{N_i \times d}$8 and $f_i \in \mathbb{R}^{N_i \times d}$9. The implementation uses RGB $f_m \in \mathbb{R}^{N_m \times d}$0 for reconstruction, LPIPS for perceptual supervision, and mask consistency between accumulated density and foreground segmentation. Optimization uses AdamW with learning rate $f_m \in \mathbb{R}^{N_m \times d}$1, batch size 64, $f_m \in \mathbb{R}^{N_m \times d}$2 supervision views, input resolution 512×512, and patch crop size $f_m \in \mathbb{R}^{N_m \times d}$3 sampled with a weighted foreground-aware policy. Training was run on 64× NVIDIA H100 GPUs.

The architecture is explicitly feed-forward. The inference pipeline is: SMPL estimation from the input image, token extraction by DINOv2 and PTv3, transformer fusion into triplane tokens, and novel-view rendering or optional mesh extraction. No diffusion stage and no per-instance NeRF/3DGS fitting are used.

4. Benchmarks, ablations, and failure modes

HumanNOVA is evaluated on THuman2, CustomHuman, and 2K2K, using both frontal and side-view inputs. Rendering metrics are PSNR, SSIM, and LPIPS, with rendered views at 512×512 and viewpoints uniformly distributed around the subject at 20° horizontal intervals. Geometry is measured with Chamfer Distance (CD, cm), Normal Consistency (NC), and F-Score (threshold 0.01 m) (Hu et al., 1 Jun 2026).

Dataset	Frontal input: PSNR / SSIM / LPIPS	Side input: PSNR / SSIM / LPIPS
CustomHuman	22.29 / 0.9360 / 42.42	22.52 / 0.9348 / 43.58
THuman2	23.96 / 0.9382 / 42.13	24.35 / 0.9382 / 42.55
2K2K	22.65 / 0.9336 / 41.72	23.07 / 0.9347 / 40.95

For frontal geometry, the paper reports CustomHuman: CD 1.062/1.102 cm, NC 0.867, F-Score 61.379; THuman2: CD 1.027/1.098 cm, NC 0.840, F-Score 61.939; 2K2K: CD 1.045/1.110 cm, NC 0.836, F-Score 60.673. Relative to SiTH, HumanNOVA reports approximately 41.8%, 37.0%, and 43.3% relative LPIPS improvements on CustomHuman, THuman2, and 2K2K under frontal input, and an approximately 94.3% relatively higher F-Score than SiTH on CustomHuman for geometry under side-view input.

The ablation results attribute performance to four factors. First, both data components are necessary: removing synthetic assets or removing real multi-camera data reduces PSNR/SSIM and increases LPIPS; the paper gives LPIPS 45.18 with both, 46.51 without assets, and 47.83 without multi-cam. Second, scale matters: LPIPS improves from 50.14 at 25% data to 47.03 at 50% data to 45.18 at full scale. Third, the mesh prior improves structural plausibility and LPIPS. Fourth, triplane capacity and fusion depth are critical; reducing triplane resolution, for example to 32, severely hurts detail and LPIPS, while reducing fusion depth from 4 to 2 causes the largest drop.

The reported limitations are also explicit. Severe ambiguity remains under occlusions and extreme poses. Dresses, overalls, loose layers, and complex garments are difficult to infer on the back side. Highly detailed hair, hats, glasses, jewelry may be simplified or misrepresented. Motion blur degrades token features and geometry, and difficult backgrounds can cause faint background leakage into density despite the mask loss. The model is described as robust to moderate SMPL inaccuracies, but grossly incorrect SMPL fits can bias structure. The paper also emphasizes ethics and privacy, noting risks of misuse, harassment, and privacy violations in realistic human reconstruction.

5. Related graphics usages: structure-aware HumanNOVA and NOVA-Human

In the HFNeRF paper, “HumanNOVA” is used in a broader sense: human-centric novel view synthesis with explicit, structure-aware human modeling. HFNeRF is introduced as a generalizable human feature NeRF that aims not only at photorealistic rendering but also at biomechanic structure, specifically skeletons and joints, for downstream AR/VR tasks. It leverages a pre-trained ResNet or DINO ViT image encoder, lifts multi-view 2D features into a 3D field through neural rendering, and predicts both photometric outputs $f_m \in \mathbb{R}^{N_m \times d}$4 and per-joint feature logits that are volume-rendered into multi-channel 2D heatmaps. The system is described as fully differentiable, jointly learning color, geometry, and human skeleton. Training uses RenderPeople multi-view sequences with 34 cameras for training and 2 held-out cameras for testing, optimized in PyTorch on an RTX 3090 with Adam for 100,000 iterations, using $f_m \in \mathbb{R}^{N_m \times d}$5 and heatmap supervision distilled from OpenPose. On RenderPeople, ResNet features yield PSNR 46.42, SSIM 0.9996, LPIPS 0.0024, and heatmap MSE 0.0003, whereas DINO features yield PSNR 35.93, SSIM 0.9914, LPIPS 0.0345, and heatmap MSE 0.0001, indicating a reported trade-off between photometric fidelity and structural feature quality (Dey et al., 2024).

A different but terminologically adjacent case appears in NOVA-3D. That paper states explicitly that HumanNOVA is not introduced as a separate concept and that, when readers say “HumanNOVA,” they are referring to the NOVA-Human dataset. NOVA-Human is a benchmark for full-body anime character reconstruction from non-overlapped front and back views. It contains 10.2k distinct 3D anime character VRM models, 20 images per character—16 randomly sampled perspective views plus 4 fixed orthogonal views—for a total of 163.2k images, all at 512×512, with accurate camera parameters. The associated NOVA-3D framework is a 3D-GAN-based method with a tri-plane representation, a dual-viewpoint encoder using ResNet for the front view and PAT (Pooling Attention Transformer) for the back view, and a direction-aware attention module that fuses front and back features according to ray direction. On the animation head reconstruction task, it improves the state-of-the-art baseline to 94.453 in SSIM, 7.726 in LPIPS, and 19.575 in PSNR on average; on full-body ablations over right, left, front, and back views, removing the direction-aware attention or the dual-viewpoint encoder degrades performance and reintroduces artifacts such as the reported “ghost face” in back views (Wang et al., 2024).

These two usages situate “HumanNOVA” within a broader graphics trajectory: from photometric human novel-view synthesis toward explicit structural conditioning, and from overlapping multi-view settings toward sparse or non-overlapped input regimes.

6. Other overloaded meanings outside 3D graphics

In comparative genomics, “HumanNOVA” is used as a distilled label for human de novo gene emergence and characterization. The underlying study analyzes strand-specific polyA+ RNA-Seq from human, chimpanzee, macaque, and mouse, combined with additional vertebrate datasets, and operationally defines de novo genes as newly expressed, multiexonic transcriptional loci lacking detectable homologs in other species. The final counts are 634 human-specific genes (1,029 transcripts), 780 chimpanzee-specific genes (1,307 transcripts), and 1,300 hominoid-specific genes (3,062 transcripts), for a total of 2,714 de novo genes expressing 5,398 transcripts, representing about 4% of all expressed multiexonic genes in human and chimpanzee. The data link de novo transcription to gains of upstream CREB, JUN (AP-1), RFX, and BREd motifs, to downstream enrichment of U1 snRNP 5′ splice site motifs, and partly to LTR contributions near transcription start sites. Translation evidence is reported for 22 de novo genes (31 transcripts) through proteomics and ribosome profiling, and translated de novo ORFs show 43.51 substitutions/Kb versus 59.65 substitutions/Kb in introns, consistent with purifying selection on that subset. The study further reports a strong testis bias, with 93.8–94.5% of de novo transcripts expressed in testis (Ruiz-Orera et al., 2015).

In AI theory, “HumanNOVA” denotes a NOVA-based blueprint for human-amplified knowledge discovery. The framework models iterative discovery as the loop generate, verify, accumulate, retrain over a valid knowledge set $f_m \in \mathbb{R}^{N_m \times d}$6 and candidate space $f_m \in \mathbb{R}^{N_m \times d}$7, with the key quantity

$f_m \in \mathbb{R}^{N_m \times d}$8

the model mass on historically undiscovered valid artifacts. For finite $f_m \in \mathbb{R}^{N_m \times d}$9, the paper gives sufficient conditions for almost-sure coverage: monotone accumulation, persistent pre-discovery exposure, artifact-wise nondegenerate acceptance, and no false positives. When these fail, the framework identifies forgetting, exploration failure, acceptance failure, and contamination. Under imperfect verification, the paper derives a contamination-to-genuine ratio

$T \in \mathbb{R}^{3hw \times d}$0

and argues that as $T \in \mathbb{R}^{3hw \times d}$1 shrinks near the frontier, even small fixed false-positive rates become unsafe. It also states that Good–Turing is only a local batch-diversity diagnostic, not an estimator of historically undiscovered valid mass. Under a tail-equivalence assumption to a Zipf law with exponent $T \in \mathbb{R}^{3hw \times d}$2, cumulative generation cost scales as

$T \in \mathbb{R}^{3hw \times d}$3

which formalizes superlinear discovery cost and motivates human intervention through guidance, generation, and verification, especially near autonomous exploration barriers (Avestimehr et al., 12 May 2026).

Across these domains, the shared label does not denote a single unified theory or artifact. Rather, it marks a convergence of interests around structured discovery under sparsity or ambiguity: single-image human avatar reconstruction in 3D vision, structure-aware view synthesis and anime reconstruction in graphics, transcriptional novelty in genomics, and human-in-the-loop frontier discovery in AI theory.