Dream3DAvatar: 3D Avatar Reconstruction

• Dream3DAvatar is a research topic encompassing pipelines that reconstruct high-fidelity, animatable 3D avatars from limited inputs using text-driven diffusion and multi-view synthesis.
• It integrates adapter-enhanced multi-view generation with methods like 3D Gaussian lifting and pose-aware UV mapping to achieve realistic geometric and texture details.
• The approach addresses challenges of monocular input and view inconsistency, demonstrating advances in identity preservation and real-time 3D avatar reconstruction.

Dream3DAvatar denotes a line of 3D avatar research concerned with reconstructing or generating high-fidelity avatars from severely limited input, and it is also the title of the paper "Dream3DAvatar: Text-Controlled 3D Avatar Reconstruction from a Single Image" (Liu et al., 16 Sep 2025). In the literature associated with the name, the problem setting ranges from generating a high fidelity 3D face avatar with a high-resolution UV texture map from a single image (Wang et al., 2019), to generating realistic animatable 3D human avatar models solely from textual descriptions (Kolotouros et al., 2023), to reconstructing a high-fidelity, animatable 3D human avatar from a single RGB image by combining multi-view generation, 3D Gaussian lifting, and pose-aware UV-space mapping (Bühler et al., 21 Jul 2025). The term therefore refers less to a single fixed architecture than to a family of pipelines that couple strong priors, view synthesis or diffusion guidance, and animation-aware 3D representations.

1. Nomenclature and conceptual scope

The most specific use of the term is the 2025 system "Dream3DAvatar: Text-Controlled 3D Avatar Reconstruction from a Single Image," which is described as an efficient and text-controllable two-stage framework for 3D avatar generation from a single image (Liu et al., 16 Sep 2025). Its first stage performs adapter-enhanced multi-view generation, and its second stage reconstructs high-fidelity 3D Gaussian Splat representations from the generated images. The paper emphasizes realistic, animation-ready 3D avatars without any post-processing.

At the same time, several adjacent papers use the label in a broader or shorthand sense. The detailed implementation guide attached to "Digital Twin: Acquiring High-Fidelity 3D Avatar from a Single Image" presents a "Dream3DAvatar pipeline" with synthetic training data generation, deep shape regression, non-rigid deformation refinement, and UV texture mapping and completion (Wang et al., 2019). The detailed overview of DreamHuman notes that the method is "sometimes referred to in shorthand as Dream3DAvatar" (Kolotouros et al., 2023). The paper "Dream, Lift, Animate: From Single Images to Animatable Gaussian Avatars" explicitly identifies its pipeline as Dream3DAvatar and defines it through multi-view dreaming, Gaussian lifting, and UV-space Gaussian decoding (Bühler et al., 21 Jul 2025).

A recurrent misconception is that Dream3DAvatar implies one canonical 3D representation. The cited literature instead spans low-poly head meshes with high-resolution UV textures (Wang et al., 2019), deformable NeRFs driven by statistical body models (Kolotouros et al., 2023), dual-observation-space NeRFs with SMPL guidance (Cao et al., 2023), DensePose-conditioned explicit voxel or mesh synthesis (Zhang et al., 2023), primitive-based 3D Gaussian avatars with implicit mesh learning (Yuan et al., 2023), and feedforward 3D Gaussian Splat reconstruction from synthesized multi-view images (Liu et al., 16 Sep 2025). This suggests that the unifying idea is not representation-specific; it is the combination of underconstrained input with increasingly explicit geometric, semantic, and identity priors.

2. Single-image head-avatar reconstruction as an early foundation

An early formulation of the Dream3DAvatar idea appears in the single-image face-avatar pipeline described for "Digital Twin" (Wang et al., 2019). The pipeline consists of four major stages: synthetic training data generation, deep shape regression, non-rigid deformation refinement, and UV texture mapping and completion. The synthetic stage begins with a small but high-quality set of neutral 3D facial scans, aligns each scan to a generic low-poly head mesh of approximately $2.9\text{K}$ vertices using non-rigid registration, and transfers the high-res diffuse texture into the mesh’s UV space at, for example, $2048 \times 2048$ resolution. Shape augmentation is performed in a deformation representation $D \in \mathbb{R}^{|P|\times 9}$, with new codes sampled by randomly combining $m=5$ neighbors and then decoded by solving a sparse linear system. Rendering uses a production renderer with $30$ pre-defined light setups, random head rotations in yaw, pitch, and roll within $[-15^\circ,15^\circ]$, and random indoor or outdoor back-plates, producing approximately $10\,000$ image–ground-truth-mesh pairs.

The coarse geometry predictor uses LightCNN-29V2 pretrained on millions of faces as an encoder that produces a $256$-D identity feature, followed by three fully connected layers of sizes $[128, 200, 3 \times 2925]$. The final fully connected layer is initialized with a PCA basis of $200$ components computed from the $2048 \times 2048$0 augmented shapes. The regression objective is a weighted per-vertex $2048 \times 2048$1 loss,

$2048 \times 2048$2

with $2048 \times 2048$3 if $2048 \times 2048$4, else $2048 \times 2048$5. The guide also notes an optional identity feature consistency term,

$2048 \times 2048$6

Training uses Adam with learning rate $2048 \times 2048$7, $2048 \times 2048$8, $2048 \times 2048$9, batch size $D \in \mathbb{R}^{|P|\times 9}$0, and $D \in \mathbb{R}^{|P|\times 9}$1 epochs on $D \in \mathbb{R}^{|P|\times 9}$2 synthetic images.

The non-rigid refinement stage augments the coarse mesh $D \in \mathbb{R}^{|P|\times 9}$3 with FaceWarehouse blendshapes, per-vertex correction $D \in \mathbb{R}^{|P|\times 9}$4, expression weights $D \in \mathbb{R}^{|P|\times 9}$5, and camera parameters $D \in \mathbb{R}^{|P|\times 9}$6. The composite mesh is

$D \in \mathbb{R}^{|P|\times 9}$7

Optimization minimizes

$D \in \mathbb{R}^{|P|\times 9}$8

where $D \in \mathbb{R}^{|P|\times 9}$9 is the landmark alignment term, $m=5$0 is the corrective field regularization term, and $m=5$1 encodes expression and camera priors. Initialization sets $m=5$2, $m=5$3, obtains $m=5$4 from EPnP on initial landmarks, sets $m=5$5, and solves the objective by $m=5$6 Gauss–Newton iterations with typical weights $m=5$7, $m=5$8, $m=5$9, $30$0, and $30$1.

The UV stage projects visible texels into UV space using the refined mesh and fills unseen regions by identifying the closest subject in identity-feature space using $30$2 distance on LightCNN embeddings, borrowing their full-head UV, and blending with Poisson image editing. The reported runtime is approximately $30$3–$30$4 ms per image: shape-regression inference at approximately $30$5 ms on GPU, non-rigid refinement at approximately $30$6 ms, and texture projection plus Poisson blending below $30$7 ms. The system is described as real-time on desktop and mobile-friendly via optimized OpenGL ES.

3. Diffusion-guided full-body avatar generation and the emergence of body priors

Later Dream3DAvatar-related systems shift from facial reconstruction to full-body, animatable human avatars, typically by coupling diffusion supervision to parametric body priors. DreamHuman defines the task as generating realistic animatable 3D human avatar models solely from textual descriptions, and combines a deformable NeRF with an imGHUM/SMPL-style implicit human model that provides signed distance and semantic correspondence codes (Kolotouros et al., 2023). Its final density is

$30$8

with the body-density proxy $30$9, and optimization includes Score Distillation Sampling, density, normal, mask, orientation, proposal, shape, and skinning-related regularizers. DreamHuman also introduces semantic zoom supervision through six regional crops and reports CLIP R-Precision improvements over DreamFusion, with $[-15^\circ,15^\circ]$0 R-Prec, Top-3 $[-15^\circ,15^\circ]$1, Top-5 $[-15^\circ,15^\circ]$2, and FID approximately $[-15^\circ,15^\circ]$3 versus DreamFusion approximately $[-15^\circ,15^\circ]$4.

DreamWaltz extends this line through 3D-consistent occlusion-aware SDS with skeleton conditioning and an explicitly animatable second stage (Huang et al., 2023). In Stage I, a canonical-pose NeRF is optimized under Stable Diffusion v1.5 plus ControlNet skeleton conditioning, with occlusion culling of invisible joints to suppress multi-face artifacts. In Stage II, inverse skinning maps posed-space points to canonical space, and a density weighting network computes

$[-15^\circ,15^\circ]$5

which reweights the NeRF density to suppress stray densities for complex non-rigged geometry such as capes or skirts. Training uses Adam with learning rate $[-15^\circ,15^\circ]$6, $[-15^\circ,15^\circ]$7 iterations for Stage I and $[-15^\circ,15^\circ]$8 for Stage II. In a user study with $[-15^\circ,15^\circ]$9 raters and $10\,000$0 prompts, DreamWaltz outperforms DreamAvatar by $10\,000$1 in geometry and $10\,000$2 in texture.

DreamAvatar makes the body prior more explicit by combining a trainable NeRF, a frozen Stable Diffusion model with LoRA adapters, and SMPL-guided shape and pose control (Cao et al., 2023). Its distinctive device is a dual-observation-space design with canonical space $10\,000$3 and posed space $10\,000$4 linked by a learnable deformation field. Both branches receive Variational Score Distillation supervision, while an additional zoomed-in head VSD loss uses landmark-conditioned ControlNet and a learned <back-view> token to resolve the multi-face “Janus” collapse and sharpen facial detail. Optimization runs for $10\,000$5 iterations per prompt on a single NVIDIA A40 GPU, with learning rates $10\,000$6 for the NeRF, $10\,000$7 for the deformation MLP, and $10\,000$8 for LoRA adapters.

AvatarVerse addresses similar failure modes through DensePose-conditioned 2D diffusion and progressive explicit 3D synthesis (Zhang et al., 2023). It renders a per-pixel part-label map from SMPL, injects the encoded DensePose signal into ControlNet’s U-Net, and replaces standard SDS with a DensePose-conditioned score function. A progressive voxel-grid schedule doubles resolution at iterations $10\,000$9, $256$0, and $256$1, shrinks the bounding box at iteration $256$2, and reduces camera radius at iterations $256$3 and $256$4. The paper reports that against DreamFusion, DreamAvatar, and DreamWaltz, AvatarVerse was preferred more than $256$5 of the time in a user study, and won $256$6 of votes against DreamHuman for “most realistic human.”

GAvatar replaces NeRF with primitive-based 3D Gaussian splatting while preserving explicit animatability (Yuan et al., 2023). Gaussians are grouped into pose-driven primitives with translation, rotation, and scale derived from SMPL and pose-dependent corrective offsets. Rather than storing Gaussian attributes independently, GAvatar predicts them through tiny implicit fields queried in a rest pose. An SDF-based implicit mesh learning module ties opacity to a zero-level set surface and extracts a detailed textured mesh through DMTet. The combination of text-driven SDS, local position regularization, Eikonal loss, silhouette consistency, and normal SDS yields animatable full-body avatars rendered at approximately $256$7 fps at $256$8 resolution on a single RTX3090Ti.

4. Alternative personalization paradigms: art portraits, albums, image guidance, and agent loops

Not all Dream3DAvatar-related work is purely text-driven. AniArtAvatar addresses one-shot 3D art-avatar reconstruction and animation from a single image (Li, 2024). Its first stage uses a view-conditioned 2D diffusion model, identified as Wonder3D, to synthesize multi-view images and normals from a single art portrait with a neutral expression. Its second stage reconstructs a static avatar with an SDF-based neural surface using NeuS and the objective

$256$9

Animation is driven by $[128, 200, 3 \times 2925]$0 landmarks reduced to $[128, 200, 3 \times 2925]$1 relevant points, which are projected to 3D via ray–SDF collision and deformed through Delaunay triangulation, while head and torso are controlled by two cages. The reported setup uses a single NVIDIA RTX 3090, six $[128, 200, 3 \times 2925]$2 views, $[128, 200, 3 \times 2925]$3 NeuS iterations, a Marching Cubes grid of $[128, 200, 3 \times 2925]$4, and achieves FID $[128, 200, 3 \times 2925]$5 and CPBD $[128, 200, 3 \times 2925]$6.

PuzzleAvatar defines the "Album2Human" task: given a personal OOTD photo collection with arbitrary articulation, partial or truncated crop, self-occlusion, and unknown camera parameters, reconstruct a faithful textured 3D avatar in canonical A-pose (Xiu et al., 2024). Its PuzzleBooth stage segments each photo into semantic assets such as face, hair, shirt, pants, and bag using Grounded SAM plus GPT-4V, assigns learned tokens, and fine-tunes a diffusion model in two stages. The 3D stage uses NFSD to optimize a deformable tetrahedral mesh initialized from an A-posed SMPL-X body. The benchmark PuzzleIOI contains $[128, 200, 3 \times 2925]$7 individuals and approximately $[128, 200, 3 \times 2925]$8 distinct subject-plus-outfit combinations. On these $[128, 200, 3 \times 2925]$9 OOTD configurations, PuzzleAvatar improves PSNR and LPIPS over TeCH and MVDreamBooth backbones, while emphasizing front-back consistency and fewer non-human artifacts.

Guide3D proposes a zero-shot text-and-image-guided pipeline based on sparse-view image generation and joint optimization of multi-resolution differentiable marching tetrahedral grids with pixel-aligned image features (Cao et al., 2023). Rather than relying only on SDS, it introduces an image-level loss using denoised renderings, a novel-view consistency loss, a boundary HED loss, and Eikonal regularization. The system uses grids at $200$0 and $200$1, diffusion sampling with $200$2 steps, guidance scale $200$3, and $200$4 training iterations. Its emphasis is direct 2D-to-3D attribute transfer through similarity-aware multi-view feature fusion.

SmartAvatar departs further from diffusion-centric optimization by introducing a vision-language-agent-driven framework for fully rigged, animation-ready 3D human avatars from a single photo or textual prompt (Huang-Menders et al., 5 Jun 2025). The system organizes generation as a four-agent loop: a Descriptor Agent parses text or image into an attribute dictionary, a Generator Agent emits Blender–HumGen3D Python calls, an Evaluator Agent computes facial similarity, anatomical plausibility, and prompt-alignment scores, and a Refiner Agent proposes delta-adjustments. Facial similarity is defined by cosine similarity between face embeddings, prompt alignment by CLIP cosine similarity, and anatomical plausibility through a bone-length penalty. The framework reports ArcFace ID similarity and CLIP_image comparisons against PSHuman and CharacterGen, and emphasizes consumer-grade hardware and animation readiness by design.

5. Dream, Lift, Animate: single-image animatable Gaussian avatars

"Dream, Lift, Animate" formulates Dream3DAvatar as a three-part pipeline from a single RGB image to an animatable Gaussian avatar (Bühler et al., 21 Jul 2025). The first part, multi-view dreaming, takes the input image $200$5 and its estimated SMPL-X pose $200$6, renders 2D skeletal control maps from virtual cameras around a $200$7 azimuth, and uses a ControlNet-guided video diffusion model to hallucinate novel views $200$8 while maintaining human appearance fidelity. The diffusion model is trained with the simplified score-matching loss

$200$9

The second part, 3D Gaussian lifting, uses a U-Net with cross-view self-attention, identified as a Large Gaussian Model architecture, to predict per-pixel Gaussian parameters from the synthesized views and their known cameras. For each Gaussian $2048 \times 2048$00, the representation includes center $2048 \times 2048$01, covariance $2048 \times 2048$02, color $2048 \times 2048$03, and opacity $2048 \times 2048$04. After confidence-based filtering, the method subsamples to approximately $2048 \times 2048$05 Gaussians.

The third part maps the unstructured Gaussian cloud into a structured UV-space latent using a transformer-based encoder. Per-Gaussian feature vectors are cross-attended against a UV-space map $2048 \times 2048$06 obtained by rasterizing the posed SMPL-X mesh. The resulting structured latent code $2048 \times 2048$07 is decoded by a SPADE-style ConvNet into a canonical map $2048 \times 2048$08 and an offset map $2048 \times 2048$09, which are fused as $2048 \times 2048$10. Given a target pose $2048 \times 2048$11 and SMPL skinning, each Gaussian is transformed by tangent-space frames obtained via linear blend skinning.

Rendering uses differentiable Gaussian splatting. Each Gaussian contributes

$2048 \times 2048$12

and volumetric compositing is approximated by front-to-back alpha compositing. The Gaussian Parameter Decoder is trained with

$2048 \times 2048$13

while the full schedule runs for $2048 \times 2048$14 epochs at $2048 \times 2048$15 and another $2048 \times 2048$16 epochs at $2048 \times 2048$17, with batch size $2048 \times 2048$18 and Adam $2048 \times 2048$19.

On ActorsHQ, the method reports LPIPS $2048 \times 2048$20, PSNR $2048 \times 2048$21, SSIM $2048 \times 2048$22 for novel view; LPIPS $2048 \times 2048$23, PSNR $2048 \times 2048$24, SSIM $2048 \times 2048$25 for novel pose; and LPIPS $2048 \times 2048$26, PSNR $2048 \times 2048$27, SSIM $2048 \times 2048$28 for joint view-plus-pose. On 4D-Dress, it reports LPIPS $2048 \times 2048$29, PSNR $2048 \times 2048$30, SSIM $2048 \times 2048$31 for novel view; LPIPS $2048 \times 2048$32, PSNR $2048 \times 2048$33, SSIM $2048 \times 2048$34 for novel pose; and LPIPS $2048 \times 2048$35, PSNR $2048 \times 2048$36, SSIM $2048 \times 2048$37 for view-plus-pose. Ablations show that direct mesh unprojection degrades LPIPS by $2048 \times 2048$38, replacing the SMPL-vertex query with a learned static query raises LPIPS to $2048 \times 2048$39, and removing pose or view conditioning in the Gaussian Parameter Decoder raises LPIPS to $2048 \times 2048$40.

6. Dream3DAvatar as a two-stage text-controllable 3D Gaussian Splat pipeline

The paper explicitly titled "Dream3DAvatar: Text-Controlled 3D Avatar Reconstruction from a Single Image" formulates the problem as full-body 3D avatar reconstruction from a single image under the additional requirement that geometry and texture of occluded regions remain text-controllable (Liu et al., 16 Sep 2025). It attributes the difficulty to the fundamentally ill-posedness of monocular input and responds with a two-stage pipeline.

Stage 1 is adapter-enhanced multi-view generation. The input is a single RGB portrait $2048 \times 2048$41 plus derived SMPL-X and skeleton renderings, and the output is $2048 \times 2048$42 pose-consistent, identity-preserving images $2048 \times 2048$43 under new viewpoints, together with text prompts $2048 \times 2048$44 for occluded regions. The Pose-Adapter injects multi-view SMPL-X renders $2048 \times 2048$45 and skeleton maps $2048 \times 2048$46 into the SDXL U-Net by extracting two feature streams with light convolution-plus-self-attention heads, average-pooling across views, and summing the two modalities to obtain $2048 \times 2048$47 for encoder blocks $2048 \times 2048$48. ID-Adapter-G preserves high-frequency facial identity by combining a face encoder, a reference image encoder, global cross-attention, local cross-attention, and row-wise attention across the concatenated multi-view batch. BLIP2 is then applied to each generated image to produce captions $2048 \times 2048$49, which are embedded by SDXL’s text encoder and incorporated into the full diffusion condition $2048 \times 2048$50.

Stage 1 optimizes only the adapter parameters through the standard diffusion reconstruction loss

$2048 \times 2048$51

with $2048 \times 2048$52 from BLIP2, $2048 \times 2048$53 from the full-body reference image, $2048 \times 2048$54 from the high-resolution face crop, and $2048 \times 2048$55 from the Pose-Adapter inputs. Face-masking with $2048 \times 2048$56 probability and random drop of any condition with $2048 \times 2048$57 probability are used to enable robust class-free guidance.

Stage 2 reconstructs a 3D avatar as a set of Gaussian splats $2048 \times 2048$58, where each primitive stores 3D mean position $2048 \times 2048$59, opacity $2048 \times 2048$60, rotation $2048 \times 2048$61, scale $2048 \times 2048$62, and RGB color $2048 \times 2048$63. Projection onto the image plane uses the fast elliptical splat renderer, expressed as

$2048 \times 2048$64

The reconstruction network is a feedforward Transformer with multi-view feature fusion. Within-view spatial attention captures local geometry, cross-view attention aligns the same UV location across views, and a learned per-UV sigmoid gate

$2048 \times 2048$65

blends the two streams before regression of all Gaussian attributes. ID-Adapter-R re-injects fine facial features in the last $2048 \times 2048$66 Transformer layers through another sigmoid-gated fusion mechanism.

Stage 2 renders the predicted 3DGS under the known viewpoints and minimizes

$2048 \times 2048$67

with typical weights $2048 \times 2048$68, $2048 \times 2048$69, and $2048 \times 2048$70. The paper states that there are no adversarial or explicit identity or pose-consistency terms beyond the multi-view fusion and adapter conditioning.

The reported experimental setup uses THuman2.1, HuGe100K, and Human4DiT, with SMPL-X fitted by Multi-HMR. Stage 1 uses $2048 \times 2048$71 reference plus $2048 \times 2048$72 target views per person; Stage 2 uses $2048 \times 2048$73 random views with at least $2048 \times 2048$74 frontal view for self-supervision. Training runs on $2048 \times 2048$75 NVIDIA A800 GPUs with AdamW, learning rates $2048 \times 2048$76 for Stage 1 and $2048 \times 2048$77 for Stage 2, $2048 \times 2048$78 iterations for Stage 1, $2048 \times 2048$79 for Stage 2, and batch size $2048 \times 2048$80.

Quantitatively, on multi-view image generation for the THuman2.1 subset, the method reports MSE $2048 \times 2048$81, PSNR $2048 \times 2048$82, SSIM $2048 \times 2048$83, and LPIPS $2048 \times 2048$84, improving over SV3D, MagicMan, PSHuman, MV-Adapter, and several ablated variants. On 3D avatar reconstruction, it reports MSE $2048 \times 2048$85, PSNR $2048 \times 2048$86, and LPIPS $2048 \times 2048$87, outperforming DreamGaussian, SIFU, CRM, and IDOL. The paper characterizes the result as state-of-the-art accuracy with greatly improved identity preservation and real-time feed-forward inference in Stage 2.

7. Recurring limitations, open problems, and persistent misunderstandings

A central limitation across Dream3DAvatar systems is the underconstrained nature of monocular or otherwise incomplete input. The 2025 Dream3DAvatar paper explicitly describes single-image full-body reconstruction as fundamentally ill-posedness because monocular input provides limited information about occluded regions (Liu et al., 16 Sep 2025). The earlier single-image head pipeline responds to a related problem through high-fidelity synthetic images intended to reduce domain gap and through identity-focused LightCNN features that focus on identity rather than lighting or pose (Wang et al., 2019). This suggests that controllability over unseen geometry and texture remains inseparable from the quality of the prior used to hallucinate them.

Another persistent issue is view inconsistency, frequently manifested as the Janus problem. DreamAvatar identifies poor facial detail and the “Janus” multi-face collapse as a central obstacle and addresses it through joint optimization of full-body and zoomed-in head VSD losses (Cao et al., 2023). AvatarVerse similarly attributes multi-faced avatars to conflicting 2D SDS signals and uses DensePose-conditioned supervision plus a smoothing regularizer to enforce exact per-view part layout (Zhang et al., 2023). The single-image Dream3DAvatar paper addresses the same family of failures with Pose-Adapter, ID-Adapter-G, BLIP2-generated text prompts for occluded regions, and ID-Adapter-R in the reconstruction stage (Liu et al., 16 Sep 2025).

The literature also exposes limitations that are specific to representation choice. AniArtAvatar requires the source portrait to start in a neutral expression or pose, does not reconstruct inner mouth, teeth, or occluded regions, and notes that zero-shot image-to-3D diffusion can fail on highly abstract or thin-geometry art such as hair or open mouth (Li, 2024). PuzzleAvatar reports that SDS-only guidance can introduce identity drift or garment hallucination, that DMTet struggles with very thin geometry such as straps or thin jewelry, and that the full PuzzleBooth-plus-SDS pipeline takes approximately $2048 \times 2048$88 hours per subject (Xiu et al., 2024). DreamHuman notes that wrinkles and hair often only appear in the albedo map rather than in density, that shading and albedo can become entangled, and that SDS cost limits per-crop resolution (Kolotouros et al., 2023).

A further misconception is that “single-image” pipelines necessarily reconstruct directly from the observed image. Several systems labeled in this way actually synthesize intermediate views before 3D reconstruction. AniArtAvatar uses a view-conditioned 2D diffusion model to synthesize multi-view images from a single art portrait (Li, 2024); Dream, Lift, Animate first dreams plausible multi-views using a video diffusion model (Bühler et al., 21 Jul 2025); and Dream3DAvatar first performs adapter-enhanced multi-view generation before feedforward 3DGS reconstruction (Liu et al., 16 Sep 2025). In these systems, single-image input refers to the supervision source, not to the absence of an internal multi-view stage.

Taken together, the Dream3DAvatar corpus marks a transition from direct monocular mesh fitting toward pipelines that explicitly separate prior construction, view completion, geometric lifting, and animation-aware representation learning. A plausible implication is that future systems will continue to hybridize strong semantic priors, controllable generation modules, and render-efficient 3D representations rather than converge on a single universal architecture.