StepFun Step1X-3D integration notice

Portions of the geometry flow transformer, shape-VAE decoder, image
conditioning, image framing, sampling loop, and surface-extraction mapping in
this family are modified from:

  stepfun-ai/Step1X-3D
  https://github.com/stepfun-ai/Step1X-3D
  revision cb5ac944709c6c913109070c7b90c3447f57f3d4

Those portions are provided under the accompanying Apache License 2.0
(LICENSE-APACHE-2.0). Copyright (c) StepFun. All ported source files carry a
prominent modification notice. Independently written diffusers-3d package glue
remains under the package Apache-2.0 license.

The texture stage's attention processors, multi-view loop, and conversion
are modified from Step1X-3D's step1x3d_texture (Apache-2.0), which adapts
MV-Adapter (https://github.com/huanngzh/MV-Adapter, Apache-2.0). Its latent
texture synchronisation is modified from Step1X-3D's texture_sync, which
derives from SyncMVD (https://github.com/LIU-Yuxin/SyncMVD); those portions
are provided under the accompanying MIT License (LICENSE-MIT), Copyright (c)
2023 LIU-Yuxin. The CuPy Voronoi solver in texture_sync carries no license
and is not used.

Step1X-3D's shape VAE descends from Michelangelo
(https://github.com/NeuralCarver/Michelangelo, GPL-3.0), whose perceiver
attention layers in turn follow OpenAI Shap-E (MIT). Step1X-3D distributes its
version without a GPL notice under Apache-2.0; this family re-implements the
decoder on stock Diffusers blocks and copies no Michelangelo source.

Not included: upstream files carrying the Tencent Hunyuan 3D non-commercial
license header (step1x3d_geometry/models/autoencoders/volume_decoders.py and
the texture-stage custom_rasterizer and differentiable_renderer). Occupancy is
evaluated on a dense grid written for this package instead of upstream's
hierarchical volume decoder. pymeshlab (GPL) post-processing is not used.

The converted weights come from https://huggingface.co/stepfun-ai/Step1X-3D
(Apache-2.0; layout checked at revision
bf7084495b3a72222f36549b7942948aa4d9daa7). The visual encoder weights are a
DINOv2-with-registers ViT-L/14 (facebook/dinov2-with-registers-large,
Apache-2.0). The texture converter combines the release's adapter weights
(Step1X-3D-Texture, Apache-2.0) with stabilityai/stable-diffusion-xl-base-1.0
(CreativeML Open RAIL++-M License) and madebyollin/sdxl-vae-fp16-fix (MIT).
Model weights keep their own licenses.

No affiliation, sponsorship, or endorsement by StepFun, Meta, Tencent,
OpenAI, or Stability AI is claimed.