TripoSplat — component mirror (PoSTMEDIA)

A convenience mirror of the checkpoints SplatX Studio needs to run TripoSplat (single image → 3D Gaussians, developed by TripoAI) from one place.

Every file here belongs to its original author. Nothing in this repository is authored by PoSTMEDIA — see the per-component licences below, which travel with the files and prevail over any single tag you may see in this repo's metadata.

Built with DINOv3 — this repository redistributes Meta's DINOv3 weights under the DINOv3 License, a copy of which is included as LICENSE-DINOv3.md.

Licences, per component

This repository is not MIT as a whole. Four different licences apply:

Full texts: LICENSE-COMPONENTS.md (index) and LICENSE-DINOv3.md (Meta's agreement in full).

If you redistribute these files

• DINOv3 — you must ship a copy of Meta's agreement with the weights and prominently display "Built with DINOv3" on a related website, user interface, about page or product documentation (DINOv3 License §1(b)(i)(A)–(B)). Meta gates the original download; that gate is Meta's choice, but the agreement travels with any copy you pass on.
• FLUX.2 VAE (Apache-2.0) — keep the licence notice and state that the file was obtained from FLUX.2-klein-4B.
• MIT components — keep the copyright and permission notice.

⚠️ The VAE here is the Apache-2.0 one, deliberately

vae/flux2-vae.safetensors in this repository is Black Forest Labs' VAE as published in FLUX.2-klein-4B under Apache-2.0 (vae/diffusion_pytorch_model.safetensors, renamed to the filename TripoSplat's loader expects, 168,120,878 bytes, bf16).

It is not the copy from FLUX.2-dev, which is gated and carries the FLUX.2 Non-Commercial License. Earlier revisions of this mirror served that file; it has been removed and the repository recreated so the non-commercial weights are not in its history either.

The two are numerically the same model. We verified all 251 tensors / 84,046,372 parameters and found them bit-identical after the bfloat16 cast that TripoSplat's loader performs at load time (dtype=torch.bfloat16), so this substitution changes nothing about the output — an end-to-end A/B on the same input, with the sampler RNG pinned, produced a byte-identical result.

⚠️ If you fetch this VAE yourself, take it from FLUX.2-klein-4B. FLUX.2-klein-9B is a different release and is non-commercial.

Contents

📂 TripoSplat/
├── 📂 background_removal/
│   └── birefnet.safetensors                    444 MB   MIT
├── 📂 clip_vision/
│   └── dino_v3_vit_h.safetensors              1.68 GB   DINOv3 License
├── 📂 diffusion_models/
│   └── triposplat_fp16.safetensors             741 MB   MIT
└── 📂 vae/
    ├── flux2-vae.safetensors                   168 MB   Apache-2.0  (FLUX.2-klein-4B)
    └── triposplat_vae_decoder_fp16.safetensors 576 MB   MIT

Usage

With the official inference code, place the files under ckpts/:

📂 TripoSplat/
└── 📂 ckpts/
    ├── 📂 background_removal/birefnet.safetensors
    ├── 📂 clip_vision/dino_v3_vit_h.safetensors
    ├── 📂 diffusion_models/triposplat_fp16.safetensors
    └── 📂 vae/
        ├── flux2-vae.safetensors
        └── triposplat_vae_decoder_fp16.safetensors

With ComfyUI, the same tree under ComfyUI/models/.

Links

• Model & code: VAST-AI-Research/TripoSplat · paper · technical blog
• Upstream weights: VAST-AI/TripoSplat