Skip to content
← Models
Open weightsIndependent researchers (Liang, Yang, Lin, Li, Xu, Chen)/10 trustMetrics-checked

LucidDreamer

High-fidelity text-to-3D via Interval Score Matching — a fix for the over-smoothing of SDS.

Visit official page →

About

LucidDreamer, a CVPR 2024 paper, starts from a diagnosis: Score Distillation Sampling, the workhorse of text-to-3D, gives inconsistent low-quality update directions that over-smooth the result. The fix is Interval Score Matching — deterministic diffusing trajectories plus interval-based score matching — combined with 3D Gaussian Splatting in the generation pipeline. The authors report that it substantially beats the state of the art in both quality and training efficiency, at the cost of about 35 minutes on an A100 per asset.

The training code is public on GitHub, though the repo carries no license file, which legally means all rights reserved — something to resolve before building on it. A Hugging Face Gradio demo was under development at last check. The quality claims come from the paper's own comparisons; there is no independent benchmark, and 35 minutes per asset makes it a research tool rather than a production pipeline.

Specs

Context window

N/A - 3D generation model; text prompt input

Pricing

Free (open code; no license file — confirm rights before reuse)

Free quota

Unlimited (self-hosted)

API access

No hosted API; self-host the open code; Hugging Face demo was under development

Weights

Open weights

Released

Not disclosed

Benchmarks

No independent benchmarks; quality/efficiency claims are from the paper's own comparisons (author-reported)

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

text-to-3dgaussian-splattingresearchhigh-fidelity-assets

Trust score — 7/10

/10 trustMetrics-checked

Scored 7/10 (metrics tier, metrics-v1). Evidence: CVPR 2024 paper; training code public on GitHub; cited as a reference method in the text-to-3D literature. Strengths: peer-reviewed, a genuinely influential fix to SDS, code available. Docked for: (1) no license file in the repo — legal status of reuse is unclear; (2) no independent benchmarks, claims are author-reported; (3) ~35 min per asset on A100 is impractical for production. License: no license declared; treat as all rights reserved until clarified.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
6.0/10
Maintenance · recency, releases, issue load
4.0/10
Openness · license + docs
6.0/10
Manual review · sanity read + red-flag check
8.0/10
  • · CVPR 2024 paper (Interval Score Matching)
  • · Training code public; no license file in repo
  • · ~35 min per asset on A100 (author-reported)
  • · No independent benchmarks

Sources