mlx-community/tripo_splat_mlx

TripoSplat MLX

MLX-native implementation of TripoSplat for Apple Silicon.

Part of TripoFlux MLX — a full MLX / CoreML accelerated text-to-3D pipeline (FLUX.2-klein-9B → BiRefNet → TripoSplat → .ply / .splat / .spz) with an R3F web UI featuring a Looking Glass holographic WebXR preview.

Demo

📺 Bilibili video: Looking Glass (LKG) WebXR holographic preview demo

This repository contains MLX ports of the TripoSplat submodules:

• mlx_models/dinov3_mlx.py — DinoV3 ViT-H/16+ image encoder
• mlx_models/flux2vae_mlx.py — Flux2 VAE encoder
• mlx_models/flow_mlx.py — LatentSeqMMFlowModel (24-block multimodal flow transformer)

The original TripoSplat decoder (OctreeGaussianDecoder) remains on PyTorch MPS due to its complex dynamic sampling logic.

Weights

File	Size	Description
dino_v3_vit_h.safetensors	1.6 GB	DinoV3 ViT-H/16+ image encoder
flux2-vae.safetensors	336 MB	Flux2 VAE encoder
triposplat_fp16.safetensors	741 MB	LatentSeqMMFlowModel (flow transformer)
triposplat_vae_decoder_fp16.safetensors	576 MB	TripoSplat VAE decoder
birefnet.safetensors	444 MB	BiRefNet background removal

License

This is a mixed-license repository. The most restrictive artifact determines what the repository as a whole can be used for — with flux2-vae.safetensors included, the repo as a whole is non-commercial only.

⛔ Non-commercial — separate restriction

File	Upstream	License	Terms
flux2-vae.safetensors	black-forest-labs/FLUX.2-klein-9B	FLUX Non-Commercial License	Weights are gated upstream. Commercial use of the weights or their outputs requires a paid license from Black Forest Labs. If you need a fully permissive stack, delete this file and load the VAE from the upstream gated repo at runtime after accepting its terms.

⚠️ Custom permissive license — attribution required on redistribution

File	Upstream	License	Terms
dino_v3_vit_h.safetensors	facebook/dinov3	DINOv3 License	Commercial use is allowed, but any redistribution of copies or derivatives must include the DINOv3 License text. Upstream downloads are gated and unavailable in sanctioned jurisdictions.

✅ MIT — permissive core

File	Upstream	License
triposplat_fp16.safetensors	VAST-AI/TripoSplat	MIT
triposplat_vae_decoder_fp16.safetensors	VAST-AI/TripoSplat	MIT
birefnet.safetensors	ZhengPeng7/BiRefNet	MIT
mlx_models/*.py	this repo (MLX port)	MIT

The MLX port code in mlx_models/ is released under the MIT License, same as the main project: https://github.com/Jup33Q/tripoflux-mlx.

Usage

from mlx_models.dinov3_mlx import DinoV3ViT
from mlx_models.flux2vae_mlx import Flux2VAEEncoder
from mlx_models.flow_mlx import LatentSeqMMFlowModel

# Load models
dino = DinoV3ViT()
dino.load_safetensors("dino_v3_vit_h.safetensors")

vae = Flux2VAEEncoder()
vae.load_safetensors("flux2-vae.safetensors")

flow = LatentSeqMMFlowModel(
    q_token_length=8192, in_channels=16, cam_channels=5, out_channels=16,
    model_channels=1024, cond_channels=1280, cond2_channels=128,
    num_refiner_blocks=2, num_blocks=24, num_heads=16, mlp_ratio=4,
    qk_rms_norm=True, share_mod=True, use_shift_table=True,
)
flow.load_safetensors("triposplat_fp16.safetensors")

Full Pipeline

For the complete text-to-3D pipeline (FLUX.2-klein-9B → BiRefNet → TripoSplat → SPZ export), see the main project: https://github.com/Jup33Q/tripoflux-mlx

• R3F web UI with splat viewport, SSE progress streaming, and Looking Glass WebXR preview
• SPZ-canonical exports (.spz / .ply / .splat)
• TripoSplat sampler steps selectable (20 / 24 / 28)
• Mirror on ModelScope: Jup33Q/TripoSplat-MLX