VAST-AI/TripoSplat (MIT) — single image →
3D Gaussian splats (.ply/.splat), the zoo’s first 3D model. Outputs drop straight
into a Gaussian-splat viewer (RealityKit on visionOS, or
MetalSplatter on iOS/macOS).
Bundle: 🤗 mlboydaisuke/TripoSplat-CoreAI
— 5 nets converted (each gated converted-vs-eager cos = 1.000000): DINOv3 ViT-H
encoder + Flux2-VAE encoder (fp16), 20-step flow-matching DiT denoiser (fp16), octree
probability decoder + Gaussian decode with .ply activations baked (fp32). The
flow-matching sampler and the octree systematic resampling stay host-side (data-dependent
control flow). ~1 min per image on a Mac GPU.
▶️ Run it (source) — apps/TripoSplatMac,
the zoo app that ships this model (single image → 3D Gaussian splats on Mac: 5 converted nets + host flow sampler and octree resampling; build & run steps in its README).
image ──(bg removal, host)──▶ 1024²
DINOv3 ViT-H (1,3,1024,1024) → (1,4101,1280) feat1
Flux2-VAE enc (1,3,1024,1024) → (1,4096,128) feat2
DiT ×20 steps: latent(1,8192,16) + cam + t + feat1 + feat2 → latent (host FlowEuler sampler)
octree decoder: positions + cond → occupancy logits (host systematic resampling)
decode: points + cond → (262144,14) splats → .ply / .splat
Conversion + runner scripts: conversion/triposplat/.
The zoo’s TripoSplatMac app: drop an image, get a .ply/.splat
you can open in a splat viewer. (The 5-net + host-sampler pipeline is an engine showcase —
a kit threeD surface is future work.)