Research map / 3D vision / Gaussian splatting
Arc2Avatar: Generating Expressive 3D Avatars from a Single Image via ID Guidance
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Arc2Avatar generates expressive, blendshape-drivable 3D Gaussian Splat head avatars from a single image by distilling a LoRA-extended Arc2Face face foundation model via ISM-based SDS with FLAME-anchored splats.
Problem
Generating realistic, identity-preserving, expressive 3D head avatars from a single image using SDS, which typically suffers from oversaturated colors, weak identity guidance from text prompts, and lack of animation-ready correspondence.
Approach
Use an ID-conditioned face foundation model (Arc2Face) instead of text-to-image SD as the SDS guidance, extended to 360° views via LoRA fine-tuning on PanoHead synthetic renders; splats are initialized on FLAME vertices, masked from densification/pruning in the face region, and held near the template by L2 and Laplacian regularizers so blendshapes can drive the avatar.
Key results
- Arc2Avatar achieves the lowest FID of 143.86 versus 155.64 for ID-to-3D, the next best.
- In a 40-participant user study, Arc2Avatar received 93% of preference votes versus 7% for ID-to-3D.
- Identity similarity distribution shows highest scores and lowest variance across views compared with ID-to-3D and image-to-3D baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Arc2Avatar instead of ID-to-3D | 3D head avatar generation · Celebrity images | FID: 155.64 -> 143.86 | Helped |
| Use Arc2Avatar instead of ID-to-3D | 3D head avatar generation · User study, 10 identities per participant | preference votes: 7% -> 93% | Helped |
Surprising findings
- Image-to-3D methods reach high identity scores on frontal views but show high variance and very low scores because side/back views contain artifacts.
Limitations
- Expression stochasticity from Arc2Face can break neutral pose and template correspondence.
- Artifacts in ears and neck inherited from PanoHead.
- Splats can appear blurry near edges; base diffusion is 512x512.
Experimental setup
Datasets: PanoHead synthetic multi-view renders CoMA expression blendshapes celebrity evaluation set (45 celebrities, 20 images each)
Benchmarks: Identity Similarity Distribution (ISD) on 45 celebrities FID vs celebrity images user preference study (40 participants)
Metrics: ArcFace cosine identity similarity FID user preference %
Baselines: ID-to-3D Magic123 DreamCraft3D HumanNorm TADA DreamFace Fantasia3D
Builds on: Score Distillation Sampling Interval Score Matching Arc2Face 3D Gaussian Splatting FLAME
Closest papers by meaning
- Human-3Diffusion: Realistic Avatar Creation via Explicit 3D Consistent Diffusion Models (NeurIPS 2024 · similarity 0.89)
- RGBAvatar: Reduced Gaussian Blendshapes for Online Modeling of Head Avatars (CVPR 2025 · similarity 0.89)
- ID-to-3D: Expressive ID-guided 3D Heads via Score Distillation Sampling (NeurIPS 2024 · similarity 0.89)
- GPAvatar: High-fidelity Head Avatars by Learning Efficient Gaussian Projections (CVPR 2025 · similarity 0.89)
- En3D: An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data (CVPR 2024 · similarity 0.88)
- DiffusionAvatars: Deferred Diffusion for High-fidelity 3D Head Avatars (CVPR 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.