RFD 1089: HY-World 2.0, a quality path beside the fast one
Problem
weftspun_image_to_world (TripoSplat plus optional TRELLIS props) answers in minutes, on roughly 20 GB of VRAM, with a single-view splat blob, room-sized at best. Tencent’s HY-World 2.0 can produce a trained, multi-view-consistent, navigable 3D Gaussian world instead, at the cost of a multi-stage, multi-hour pipeline needing 17-billion- parameter models. Nothing scoped how that pipeline would plug into this project’s existing job and manifest contract.
Decision
Add hyworld2_image_to_world as a second, disabled-by-default model next to the existing fast path, not a replacement for it. One orchestrating adapter runs five pipeline stages plus an optional panorama step as subprocesses (panorama, trajectory generation, trajectory render, world expansion, Gaussian-splat data prep, and the 3DGS train itself), writing job_progress.json per stage for client polling, and only publishes the world manifest once training completes. The client needs no new task type, only a model-picker entry and a multi-stage progress UI.
See DETAILS.md for the stage table, the proposed job schema, the infrastructure prerequisites, the phased rollout, and the risk table.