Seed3D 2.0
ByteDance's simulation-ready 3D generator: single image or text in, PBR assets and full scenes out.
About
Seed3D 2.0 is ByteDance Seed team's image and text-to-3D system, released in April 2026. A coarse-to-fine diffusion transformer builds geometry while a mixture-of-experts material model produces multi-view albedo and metallic-roughness maps, outputting complete PBR assets. Beyond single objects it decomposes assets into articulated parts in URDF for simulators like Isaac Sim and can assemble full scenes from text, multi-view images, or video.
In the authors' large-scale blind study with 60 experienced 3D modelers, Seed3D 2.0 won 69 to 89.9% of textured-asset comparisons against five recent commercial models including Hunyuan3D-2.5/3.1, Tripo 3.0, and Rodin Gen2. It is served as a commercial API on Volcano Engine as Doubao-Seed3D-2.0; no open weights are available.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Author-run blind study (60 3D modelers, ~200 cases): 69.0-89.9% win rate on textured assets vs 5 commercial models (arXiv report); geometry preference 65.1-98.3% (provider-supplied, via press)
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: technical report on arXiv with detailed methodology; April 2026 release covered by 3D press; API live on Volcano Engine. Strengths: strong author-run human evals, simulation-ready outputs (URDF, PBR), scene-level generation. Docked for: (1) all benchmark wins are author-run with no independent replication; (2) closed weights, China-hosted API, no public pricing. License: proprietary commercial.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · ByteDance Seed; released 2026-04-23; coarse-to-fine DiT + MoE material model
- · Author-run blind study (60 modelers, ~200 cases): 69.0-89.9% textured win rate vs 5 commercial models
- · PBR assets, URDF part decomposition for Isaac Sim, full scene generation
- · No open weights; no public pricing; no independent replication of evals