fszontagh/gen3d-gguf

GGUF weight conversions for ggml runtimes

GGUF conversions of open image-to-3D, vision and segmentation models, for inference with ggml-based C++ runtimes rather than PyTorch.

Nothing here is a new model. Every file is a format conversion of a published checkpoint: the tensors are the upstream ones, re-laid-out and in most cases cast to F16. Conversions are verified tensor-by-tensor against the source (name, shape, dtype and payload) before publishing.

Licensing - read this first

These are derived weights, and the upstream licence travels with them. Converting a checkpoint to another container does not change its terms. Each folder below names where its weights came from; check that project's licence before using anything here, especially commercially. Several of these are not permissive:

folder	upstream	licence to check
trellis2/	Microsoft TRELLIS / TRELLIS.2	MIT
hunyuan3d-2/, hunyuan3d-2.1/	Tencent Hunyuan3D-2 / 2.1	Tencent Hunyuan Community License - has territorial and use restrictions
vision/dinov2_*	Meta DINOv2	Apache-2.0
vision/moge.gguf	Microsoft MoGe	MIT
segmentation/u2net.gguf	U^2-Net	Apache-2.0
segmentation/rmbg.gguf, rmbg2.gguf	BRIA RMBG-1.4 / 2.0	BRIA licence - non-commercial without a separate agreement
instruct-pix2pix/	InstructPix2Pix (on Stable Diffusion 1.5)	CreativeML Open RAIL-M

The table is a pointer, not legal advice, and may lag upstream changes. The authoritative terms are whatever the upstream project publishes today.

Layout

trellis2/              sparse-structure and SLAT flow models, decoders
trellis2/multiview/    the multi-view fine-tunes of the same models
hunyuan3d-2/           Hunyuan3D-2 shape DiT (+ the mv variant) and VAE
hunyuan3d-2.1/         Hunyuan3D-2.1 shape DiT and VAE
hunyuan3d-2.1/paint/   the texture-painting stack: 2.5D UNet, PBR UNet, VAE, ESRGAN
vision/                image-conditioning backbones (DINOv2, MoGe) 
segmentation/          background removal (ISNet/BRIA RMBG, BiRefNet, U^2-Net)
instruct-pix2pix/      instruction image editing (UNet + VAE)
src/                   pre-conversion sources, kept so a conversion can be redone

Notes on use

Filenames are unchanged from the flat layout. manifest.json maps role names to bare filenames, so a runtime that consumes it wants the files in one directory - the folders here are for humans. Copy what you need into a single directory and the manifest resolves as before.

Metadata. Each file carries general.architecture = gen3d and, where it is meaningful, a gen3d.component key naming the part it implements (dinov2_large, birefnet, isnet, moge_backbone, paint_esrgan, ...), plus component-specific configuration such as hidden size, layer and head counts, patch size or ESRGAN scale. Read them with any GGUF reader; nothing here depends on an out-of-band config file.

Precision. Large tensors are generally F16, with small ones (norms, biases) left F32 where the runtime expects that. A few files are F32 throughout because the source was.

src/ holds the original safetensors for a subset of the weights. They are there so a conversion can be reproduced or re-done at a different precision, not because the GGUF are lossy - conversion is a cast, and is verified.