Research map / Image generation / Text-to-image diffusion
X-Oscar: A Progressive Framework for High-quality Text-guided 3D Animatable Avatar Generation
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.
Proposes X-Oscar, a progressive Geometry→Texture→Animation framework with adaptive variational parameters and avatar-aware score distillation to generate high-quality animatable 3D avatars from text.
Problem
Text-to-3D avatar generation suffers from oversaturation and low-quality geometry and appearance, and many methods cannot produce animatable avatars.
Approach
Builds on SMPL-X with a three-stage progressive pipeline; represents avatar geometry and texture as learnable Gaussian distributions whose std is set automatically (AVP) and injects avatar-aware noise scaled by those stds into SDS (ASDS), giving an easy-to-hard learning curriculum.
Key results
- User study: X-Oscar scored 8.85 geometry quality, 8.91 texture quality, 9.22 text consistency, the highest among all compared methods.
- OpenCLIP ViT-B/32 score of 30.91 and CLIP ViT-B/32 score of 31.70 for X-Oscar, the best among compared methods on those metrics.
- Ablations show removing AVP causes geometry/appearance oversaturation and removing ASDS degrades quality (e.g., merged ears on Batman).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use X-Oscar instead of DreamWaltz | Text-guided avatar generation · User Study — Geo. Qua. | score: 6.38 -> 8.85 | Helped |
| Use X-Oscar instead of TADA | Text-guided avatar generation · User Study — Tex. Qua. | score: 6.95 -> 8.91 | Helped |
| Use X-Oscar instead of TADA | Text-guided avatar generation · User Study — Tex. Con. | score: 7.62 -> 9.22 | Helped |
| Use X-Oscar instead of TADA | Text-guided avatar generation · CLIP Score — ViT-B/32 | CLIP score: 31.09 -> 31.70 | Helped |
| Use X-Oscar instead of DreamWaltz | Text-guided avatar generation · CLIP Score — ViT-B/16 | CLIP score: 31.20 -> 31.97 | Helped |
| Use X-Oscar instead of TADA | Text-guided avatar generation · CLIP Score — ViT-L/14 | CLIP score: 27.72 -> 28.10 | Helped |
| Use X-Oscar instead of ProlificDreamer | Text-guided avatar generation · OpenCLIP Score — ViT-B/32 | OpenCLIP score: 30.81 -> 30.91 | Helped |
| Use X-Oscar instead of TADA | Text-guided avatar generation · OpenCLIP Score — ViT-B/16 | OpenCLIP score: 30.05 -> 30.28 | Helped |
| Use X-Oscar instead of Fantasia3D | Text-guided avatar generation · OpenCLIP Score — ViT-L/14 | OpenCLIP score: 31.46 -> 30.42 | Hurt |
| Use progressive modeling instead of optimizing all stages together (ablation) | X-Oscar · Avatar generation | geometry and appearance quality | Helped |
| Add AVP by representing avatars as distributions rather than specific parameters (ablation) | X-Oscar · Avatar generation | geometry and appearance realism | Helped |
| Add avatar-aware noise instead of random Gaussian noise (ablation) | X-Oscar · Avatar generation | overall geometry and appearance quality | Helped |
Limitations
- Relies on SMPL-X body prior, which limits non-human or loosely human subjects.
- Quantitative evaluation relies on a modest user study and CLIP-based metrics rather than standard benchmarks.
Experimental setup
Benchmarks: user study (40 ChatGPT-generated prompts, 52 participants)
Metrics: user study scores (geometry quality, texture quality, text c CLIP score OpenCLIP score
Baselines: DreamFusion Magic3D Fantasia3D ProlificDreamer AvatarCLIP AvatarCraft DreamWaltz HumanGaussian TADA
Builds on: DreamFusion SMPL-X Stable Diffusion VPoser
Closest papers by meaning
- AvatarVerse: High-Quality & Stable 3D Avatar Creation from Text and Pose (AAAI 2024 · similarity 0.89)
- GAvatar: Animatable 3D Gaussian Avatars with Implicit Mesh Learning (CVPR 2024 · similarity 0.87)
- DreamAvatar: Text-and-Shape Guided 3D Human Avatar Generation via Diffusion Models (CVPR 2024 · similarity 0.87)
- DiffusionGAN3D: Boosting Text-guided 3D Generation and Domain Adaptation by Combining 3D GANs and Diffusion Priors (CVPR 2024 · similarity 0.87)
- Zero-1-to-A: Zero-Shot One Image to Animatable Head Avatars Using Video Diffusion (CVPR 2025 · similarity 0.87)
- Morphable Diffusion: 3D-Consistent Diffusion for Single-image Avatar Creation (CVPR 2024 · similarity 0.86)
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.