LibreSperidlabsIdol includes code ported from Speridlabs Idol:

Source: https://github.com/speridlabs/idol-sdk (commit 4702a7f00850e34b6e83a2170cc123c3e9570744)
License: Apache-2.0.
Copyright Speridlabs (https://speridlabs.com). Authors: Alejandro Cañada, Chema Garabito.

Idol builds on TripoSplat:

Source: https://github.com/VAST-AI-Research/TripoSplat (commit d8db9e018b413dd9c4a9fe22463781bf98e8e68d)
License: MIT.
Copyright (c) 2026 VAST.

The Idol denoiser (LatentSeqMMFlowModel in nn.py) is adapted by Idol from
diffusers' transformer_wan.py:

Source: https://github.com/huggingface/diffusers
License: Apache-2.0.
Copyright 2025 The Wan Team and The HuggingFace Team.

The port covers inference: the denoiser, the octree and Gaussian decoders,
preprocessing, the image encoders' tokenization, the flow sampler and the
camera decoding. Parameter names mirror upstream so the released
safetensors load with strict=True. Changes from upstream: TensorDict inputs
are plain tensors; attention always uses PyTorch SDPA (no flash-attn
backend, no ATTN_BACKEND environment variable); randomness comes from an
explicit torch.Generator seeded like upstream's global generator; the Idol
SDK's own BiRefNet module is replaced by LibreYOLO's BiRefNet network
(libreyolo/models/birefnet, ported from ZhengPeng7/BiRefNet, MIT), which
loads the same weights; the PLY writer no longer needs plyfile; the
denoiser's cuDNN attention under compiled sampling is pinned per model
instead of through a process-wide flag, and TF32 matmuls are switched on only
while a compiled sampler runs, restoring the user's setting afterwards, with
Idol generations serialized by a process-wide lock;
grayscale-with-alpha and palette-transparency cutouts are used as cutouts
(upstream sends them to BiRefNet); shift, guidance, steps, margin and the
Gaussian count are validated before anything runs.

Weights: speridlabs/idol (Apache-2.0) and black-forest-labs/FLUX.2-klein-4B
(Apache-2.0, VAE only) are downloaded from their publishers at pinned
revisions. The DINOv3 encoder is Meta's
facebook/dinov3-vith16plus-pretrain-lvd1689m (DINOv3 License), rehosted by
LibreYOLO in bfloat16 as LibreYOLO/LibreDINOv3vith16plus together with the
DINOv3 License Agreement.

Idol citation:
Alejandro Cañada and Chema Garabito. "Idol: Single Image to 3D Gaussian
Splats." Speridlabs, 2026. https://speridlabs.com/research/idol
