Research map / 3D vision / 3D generation
RecDreamer: Consistent Text-to-3D Generation via Uniform Score Distillation
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 RecDreamer, which rectifies the diffusion prior's pose marginal to be uniform via a reweighting function and integrates it into score distillation (uniform score distillation) to mitigate the Multi-Face Janus problem in text-to-3D.
Problem
Text-to-3D score distillation produces repeated patterns across poses (Multi-Face Janus) because the pretrained text-to-image prior is biased toward canonical (frontal) poses.
Approach
Reweight the prior density with an auxiliary function so the pose marginal becomes uniform, then fold this rectified distribution into VSD-style score distillation via an extra gradient term, using a training-free DINOv2-based pose classifier and EMA estimates of the pose marginal.
Key results
- USD achieves the best FID (165.97) and uFID (165.25) versus VSD at 168.19 and 169.66 in the quantitative table.
- USD attains the highest pose entropy pEnt of 1.2488 compared with 1.0676 for VSD and 1.1542 for SDS.
- USD's CLIP score (0.6842) is slightly worse than VSD's (0.6807), which the authors attribute to back and side views not matching frontal prompt descriptions.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply uniform score distillation instead of VSD | Text-to-3D generation · Text-to-3D prompts | FID: 168.19 -> 165.97 | Helped |
| Apply uniform score distillation instead of VSD | Text-to-3D generation · Text-to-3D prompts | uFID: 169.66 -> 165.25 | Helped |
| Apply uniform score distillation instead of VSD | Text-to-3D generation · Text-to-3D prompts | pEnt: 1.0676 -> 1.2488 | Helped |
| Apply uniform score distillation instead of VSD | Text-to-3D generation · Text-to-3D prompts | CLIP: 0.6807 -> 0.6842 | Hurt |
| Apply uniform score distillation instead of VSD | Text-to-3D generation · Text-to-3D prompts | cEnt: 1.0276 -> 1.0375 | Helped |
| Directly predict pose categories from noisy images instead of using the clean-image approximation (ablation) | USD · Text-to-3D generation | Multi-Face Janus mitigation | Hurt |
| Estimate the pose distribution from pre-training samples and keep it fixed instead of updating it with EMA (ablation) | USD · Text-to-3D generation | Pose rectification | Mixed |
Surprising findings
- Reported CLIP scores are marked as lower-is-better in the table, yet the text treats USD's lower score as worse alignment, a sign of inconsistent metric direction.
Limitations
- Rectification only reweights existing modes, so it cannot generate poses or content absent from the original prior; auxiliary prompts are needed to ensure coverage.
- Visual artifacts remain and geometry is less smooth than SDS variants; texture quality is comparable to VSD rather than better.
- Evaluation uses a small set of 22 prompts, limiting statistical power of the quantitative comparison.
Experimental setup
Datasets: 22 object text prompts with auxiliary side/back descriptions
Benchmarks: 22-prompt text-to-3D comparison set
Metrics: FID uFID categorical entropy (cEnt) pose entropy (pEnt) CLIP score
Baselines: SDS SDS-Bridge VSD PerpNeg Debiased-SDS ESD
Builds on: VSD DPS DreamFusion Perp-Neg Debiased-SDS
Closest papers by meaning
- GenHMR: Generative Human Mesh Recovery (AAAI 2025 · similarity 0.87)
- 3D-SceneDreamer: Text-Driven 3D-Consistent Scene Generation (CVPR 2024 · similarity 0.86)
- TrioPose: Native Triple-Stream Diffusion Transformers for Pose-Guided Text-to-Image Generation (NeurIPS 2026 · similarity 0.86)
- TokenHMR: Advancing Human Mesh Recovery with a Tokenized Pose Representation (CVPR 2024 · similarity 0.86)
- Sculpt3D: Multi-View Consistent Text-to-3D Generation with Sparse 3D Prior (CVPR 2024 · similarity 0.86)
- DIRECT-3D: Learning Direct Text-to-3D Generation on Massive Noisy 3D Data (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.