Direct3D-S2
Gigascale text-to-3D under MIT license: 1024^3 generation with spatial sparse attention on just 8 GPUs.
About
Direct3D-S2, accepted to NeurIPS 2025, is aimed at the hardest scaling problem in 3D generation: resolution. Its trick is spatial sparse attention — an attention mechanism designed for sparse volumetric data — plus a unified sparse VAE that keeps a consistent sparse volumetric format across input, latent and output. The headline claim is training at 1024^3 resolution on only 8 GPUs, where comparable volumetric methods typically need 32+ GPUs for 256^3.
The code is on GitHub under the MIT license with about 1,300 stars and was still being pushed in late 2025, so it is genuinely usable research software rather than a paper-only artifact. The generation-quality claims are author-reported from the paper's own experiments. Like all volumetric approaches it demands serious GPU hardware to train, but inference is the practical path for most users.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Paper-reported: 1024^3 training on 8 GPUs vs 32+ GPUs typically needed for 256^3; SOTA quality/efficiency claims are author-reported
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: accepted to NeurIPS 2025; MIT-licensed repo with ~1,300 stars and recent commits; trained on public datasets. Strengths: permissive license, strong scaling story, real usable code. Docked for: (1) quality claims are author-reported, no independent benchmark yet; (2) heavy GPU requirements for training; (3) relatively new, so community validation is thin. License: MIT.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · NeurIPS 2025 accepted paper
- · MIT-licensed GitHub repo, ~1,300 stars, active through late 2025
- · Trained on public datasets
- · Efficiency claims author-reported