Research map / Image generation / Fast diffusion sampling
Consistent3D: Towards Consistent High-Fidelity Text-to-3D Generation with Deterministic Sampling Prior
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.
Consistent3D replaces SDS with a consistency distillation sampling (CDS) loss that distills ODE deterministic sampling priors from Stable Diffusion into NeRF/mesh text-to-3D models, using fixed noise and a random-interval time schedule.
Problem
Score distillation sampling (SDS) for text-to-3D generation is vulnerable to geometry collapse and poor textures, which the authors attribute to the randomness of the SDE sampling trajectory it implicitly follows.
Approach
Reinterpret SDS as SDE trajectory sampling and instead distill the deterministic ODE probability-flow trajectory: at each iteration, perturb the rendered image with fixed noise, take one Euler ODE step to a less-noisy sample, and use it as the target for the more-noisy sample (consistency-style distillation into the 3D model).
Key results
- CLIP R-Precision of 0.348 vs 0.336 for ProlificDreamer, 0.311 for Magic3D and 0.310 for DreamFusion on 40 DreamFusion-gallery prompts.
- Ablations show random timestep schedule, predetermined timestep schedule, and per-iteration random noise each degrade quality versus the full method.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CDS instead of SDS | Text-to-3D generation · DreamFusion gallery prompts | CLIP R-Precision: 0.310 -> 0.348 | Helped |
| Use CDS instead of SDS | Text-to-3D generation · DreamFusion gallery prompts | CLIP R-Precision: 0.311 -> 0.348 | Helped |
| Use CDS instead of VSD | Text-to-3D generation · DreamFusion gallery prompts | CLIP R-Precision: 0.336 -> 0.348 | Helped |
| Use a random time step schedule (ablation) | Consistent3D · Text-to-3D prompts in Figure 6 | geometry and texture quality | Hurt |
| Use a predetermined time-step schedule (ablation) | Consistent3D · Text-to-3D prompts in Figure 6 | optimization quality | Hurt |
| Use random noise in each iteration instead of fixed noise (ablation) | Consistent3D · Text-to-3D prompts in Figure 6 | convergence quality | Hurt |
Limitations
- Theoretical guarantee (Theorem 1) is an O(Delta) bound that assumes Lipschitz diffusion model and convergence, and is not validated empirically in depth.
- Evaluation is limited to 40 prompts for CLIP R-Precision plus a user study.
Experimental setup
Datasets: DreamFusion gallery prompts
Benchmarks: DreamFusion gallery prompts (40 prompts) CLIP R-Precision
Metrics: CLIP R-Precision user study preference
Baselines: DreamFusion Magic3D ProlificDreamer
Builds on: DreamFusion Consistency Models Magic3D ProlificDreamer threestudio
Closest papers by meaning
- Consistent Flow Distillation for Text-to-3D Generation (ICLR 2025 · similarity 0.94)
- Let 2D Diffusion Model Know 3D-Consistency for Robust Text-to-3D Generation (ICLR 2024 · similarity 0.91)
- Rethinking Score Distilling Sampling for 3D Editing and Generation (ICML 2025 · similarity 0.91)
- Walking the Schrödinger Bridge: A Direct Trajectory for Text-to-3D Generation (NeurIPS 2025 · similarity 0.91)
- Diffusion Time-step Curriculum for One Image to 3D Generation (CVPR 2024 · similarity 0.90)
- DreamTime: An Improved Optimization Strategy for Diffusion-Guided 3D Generation (ICLR 2024 · similarity 0.90)
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.