Research map / Image generation / Fast diffusion sampling
Advancing Text-to-3D Generation with Linearized Lookahead Variational Score Distillation
Yu Lei, Bingde Liu, Qingsong Xie, Haonan Lu, Zhijie Deng
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| update the LoRA model more steps per iteration (gamma 1 to 10) (ablation) | VSD · hamburger prompt, NeRF stage | qualitative 3D quality: gamma = 1, 2, 5, 10 | No clear effect |
| add lookahead (update LoRA before the 3D model) | VSD · DreamFusion gallery (20 prompts) | Averaged FID: 301.54 -> 496 | Hurt |
| add lookahead (update LoRA before the 3D model) | VSD · DreamFusion gallery (20 prompts) | Averaged CLIP angle: 0.324 -> 0.337 | Hurt |
| add lookahead (update LoRA before the 3D model) (ablation) | VSD · hamburger prompt, NeRF stage | speed of geometry and texture formation | Mixed |
| increase LoRA steps gamma to 2 and 5 in lookahead VSD (ablation) | L-VSD · hamburger prompt, NeRF stage | qualitative 3D quality: gamma = 2, 5 | Hurt |
| scale LoRA learning rate by 0.1 (ablation) | L-VSD · hamburger prompt, NeRF stage | qualitative 3D quality: learning rate x 0.1 | Mixed |
| add linearization of the lookahead correction (first-order term only) | L-VSD · DreamFusion gallery (20 prompts) | Averaged FID: 496 -> 284.06 | Improved |
| replace VSD with L2-VSD | VSD · DreamFusion gallery (20 prompts) | Averaged FID: 301.54 -> 284.06 | Improved |
| replace VSD with L2-VSD | VSD · DreamFusion gallery (20 prompts) | Averaged CLIP angle: 0.324 -> 0.285 | Improved |
| use high-order term instead of linear term for correction (HL-VSD) (ablation) | L2-VSD · qualitative prompts (telephone, exercise bike) | qualitative 3D quality | Hurt |
| vary lookahead scale eta, down to 1e-3 (ablation) | L2-VSD · hamburger prompt | qualitative 3D quality: eta = 1e-3 | No clear effect |
| use last-layer approximation for the correction (ablation) | L2-VSD · threestudio coarse stage | time cost (s/iteration): 1.0 -> 0.8 s/iteration; total time 2.7 -> 2.5 hrs | Mixed |
Closest papers by meaning
- Let 2D Diffusion Model Know 3D-Consistency for Robust Text-to-3D Generation (ICLR 2024 · similarity 0.90)
- Text-to-3D with Classifier Score Distillation (ICLR 2024 · similarity 0.90)
- Taming Mode Collapse in Score Distillation for Text-to-3D Generation (CVPR 2024 · similarity 0.90)
- Retrieval-Augmented Score Distillation for Text-to-3D Generation (ICML 2024 · similarity 0.90)
- Walking the Schrödinger Bridge: A Direct Trajectory for Text-to-3D Generation (NeurIPS 2025 · similarity 0.89)
- Consistent3D: Towards Consistent High-Fidelity Text-to-3D Generation with Deterministic Sampling Prior (CVPR 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.
