Research map / 3D vision / 3D generation
Consistent Flow Distillation for Text-to-3D 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 Consistent Flow Distillation (CFD), which distills 2D diffusion clean-flow ODE/SDE gradients into 3D representations using multi-view consistent Gaussian noise computed via a Noise Transport Equation warp.
Problem
Score Distillation Sampling for text-to-3D generation has maximum-likelihood-seeking behavior that degrades visual quality and diversity; existing fixes such as VSD are slow and SDS-style noise is view-independent.
Approach
Reformulate the image PF-ODE/SDE into a clean variable free of Gaussian noise, so rendered views can be used directly, and make the injected noise multi-view consistent on the object surface via a Noise Transport Equation-based warp from query views to a shared reference space.
Key results
- CFD achieves 3D-FID 78.13 versus 83.02 for VSD, 86.00 for ISM and 88.06 for SDS on 10 prompts.
- CFD achieves 3D-CLIP 35.16 versus 35.10 for VSD and 35.07 for SDS.
- Automated PickScore win rate of CFD over SDS is 0.64 and over VSD is 0.68.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace independently sampled noise with multi-view consistent Gaussian noise aligned to the object surface | CFD text-to-3D generation · Text-to-3D evaluation | 3D-FID: 83.02 -> 78.13 | Helped |
| replace independently sampled noise with multi-view consistent Gaussian noise aligned to the object surface | CFD text-to-3D generation · Text-to-3D evaluation | 3D-CLIP: 35.10$\pm$0.20 -> 35.16$\pm$0.23 | Helped |
| use gamma=0.0001 instead of gamma=0.0 (ablation) | CFD · 3D generation, gamma ablation | 3D-IS: 2.24$\pm$0.12 -> 2.60$\pm$0.21 | Helped |
| increase noise injection rate from gamma=0.0001 to gamma=0.001 (ablation) | CFD · 3D generation, gamma ablation | 3D-IS: 2.60$\pm$0.21 -> 2.47$\pm$0.39 | Hurt |
| increase noise injection rate from gamma=0.0001 to gamma=0.01 (ablation) | CFD · 3D generation, gamma ablation | 3D-IS: 2.60$\pm$0.21 -> 2.08$\pm$0.04 | Hurt |
| increase noise injection rate from gamma=0.0001 to gamma=1.0 (ablation) | CFD · 3D generation, gamma ablation | 3D-IS: 2.60$\pm$0.21 -> 1.77$\pm$0.13 | Hurt |
| use original noisy PF-ODE variable directly rather than the clean variable (ablation) | CFD text-to-3D generation · Qualitative flow-space ablation | training stability | Hurt |
| replace Noise Transport Equation noise with bilinear-interpolated noise (ablation) | CFD text-to-3D generation · Qualitative noise-design ablation | visual quality | Hurt |
| replace surface-aligned consistent noise with random noise as in SDS (ablation) | CFD text-to-3D generation · Qualitative noise-design ablation | visual quality | Hurt |
| distill only MVDream rather than using the two-stage MVDream then Stable Diffusion pipeline (ablation) | CFD complex-prompt pipeline · Complex-prompt 3D generation | visual quality | Hurt |
| use the two-stage pipeline, initializing shape with MVDream and refining with Stable Diffusion v2.1 (ablation) | CFD complex-prompt pipeline · Complex-prompt 3D generation | multi-face issues | Helped |
| add negative prompts (ablation) | CFD text-to-3D generation · 3D generation at small timesteps | sample color and style | Helped |
Surprising findings
- Directly distilling the noisy-variable PF-ODE into 3D fails after several iterations, motivating the clean variable reformulation.
- Gamma above 0.0001 produced over-smoothed textures, so only a small stochastic injection is used despite SDE often helping image generation.
Limitations
- Generation takes one to a few hours per asset.
- Multi-face Janus problem persists when distilling Stable Diffusion, and complex prompts are hard.
- Rendered views may not exactly match 2D flows due to 3D representation interference across views.
Experimental setup
Benchmarks: 3D-FID 3D-CLIP 3D-IS Aesthetics Score PickScore
Metrics: FID CLIP score Inception Score win rate gradient variance
Baselines: SDS VSD ISM FSD DreamFusion ProlificDreamer HiFA LucidDreamer
Builds on: DreamFusion ProlificDreamer Integral Noise Noise Transport Equation EDM
Closest papers by meaning
- Consistent3D: Towards Consistent High-Fidelity Text-to-3D Generation with Deterministic Sampling Prior (CVPR 2024 · similarity 0.94)
- Rethinking Score Distilling Sampling for 3D Editing and Generation (ICML 2025 · similarity 0.90)
- IM-3D: Iterative Multiview Diffusion and Reconstruction for High-Quality 3D Generation (ICML 2024 · similarity 0.89)
- Progressive Rendering Distillation: Adapting Stable Diffusion for Instant Text-to-Mesh Generation without 3D Data (CVPR 2025 · similarity 0.89)
- Let 2D Diffusion Model Know 3D-Consistency for Robust Text-to-3D Generation (ICLR 2024 · similarity 0.89)
- DreamFlow: High-quality text-to-3D generation by Approximating Probability Flow (ICLR 2024 · similarity 0.89)
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.