Skip to content
ModelListedExternal
Download

TRELLIS.2 4B GGUF

GGUF (f16) conversions of microsoft/TRELLIS.2-4B for the trellis2cpp / ggml runtime — a CUDA-free, PyTorch-free C++ port of the TRELLIS.2 image-to-3D pipeline. Load these directly; no safetensors conversion or Python...

Sourcehuggingface.co/LocalAI-io/TRELLIS.2-4B-GGUF/tree/8d55d0a0fe41c6328c1e8d8c1b3efd235defdaa0
KindOtherBase modelmicrosoft/TRELLIS.2-4BVersionRevision 8d55d0aBgrade

Model source

Kind
Other
File size
676 MB–2.44 GB
Base model
microsoft/TRELLIS.2-4B
Version
vRevision 8d55d0a
Updated
Oct 10, 2026
Downloads
12.2K
Source
Hugging Face

Source description

TRELLIS.2-4B — GGUF (f16)

GGUF (f16) conversions of microsoft/TRELLIS.2-4B for the trellis2cpp / ggml runtime — a CUDA-free, PyTorch-free C++ port of the TRELLIS.2 image-to-3D pipeline. Load these directly; no safetensors conversion or Python is needed at inference.

Files

FilePipeline stageSource safetensors
ss_flow_f16.ggufSparse-structure flow (643 occupancy)ss_flow_img_dit_1_3B_64_bf16
slat_flow_f16.ggufShape-SLAT flow, 512 fineslat_flow_img2shape_dit_1_3B_512_bf16
slat_flow_1024_f16.ggufShape-SLAT flow, 1024 cascadeslat_flow_img2shape_dit_1_3B_1024_bf16
shape_dec_f16.ggufShape decoder → dual-grid fieldsshape_dec_next_dc_f16c32_fp16
shape_enc_f16.ggufShape encoder (re-encode for texture flow)shape_enc_next_dc_f16c32_fp16
tex_dec_f16.ggufPBR texture decodertex_dec_next_dc_f16c32_fp16
tex_slat_flow_512_f16.ggufTexture-SLAT flow, 512slat_flow_imgshape2tex_dit_1_3B_512_bf16
tex_slat_flow_1024_f16.ggufTexture-SLAT flow, 1024slat_flow_imgshape2tex_dit_1_3B_1024_bf16

The image conditioning encoder (DINOv3) and the sparse-structure decoder are published separately: dinov3-vitl16-pretrain-lvd1689m-GGUF and TRELLIS-image-large-GGUF.

Conversion

Converted from the upstream bf16/fp16 safetensors to GGUF f16 (ggml ftype 1) with the convert_*_to_gguf.py scripts in the trellis2cpp project. No weights were retrained or modified — this is a format conversion only.

License

MIT, inherited from microsoft/TRELLIS.2-4B. See LICENSE. Copyright (c) Microsoft Corporation.

Sources

1 source
Hugging FacePrimaryVersion 8d55d0a0fe41c6328c1e8d8c1b3efd235defdaa0
huggingface.co/LocalAI-io/TRELLIS.2-4B-GGUF/tree/8d55d0a0fe41c6328c1e8d8c1b3efd235defdaa0

Verified Oct 10

Model artifacts

8 artifacts

shape_dec_f16.gguf

gguf · 905 MB · SHA-256 6fe53f1d7763…3530 · Hugging Face

Download

shape_enc_f16.gguf

gguf · 676 MB · SHA-256 3ec80ff58098…696c · Hugging Face

Download

slat_flow_1024_f16.gguf

gguf · 2.44 GB · SHA-256 b6a2270131e2…6bc5 · Hugging Face

Download

slat_flow_f16.gguf

gguf · 2.44 GB · SHA-256 2f94bad7b1c5…c15c · Hugging Face

Download

ss_flow_f16.gguf

gguf · 2.44 GB · SHA-256 1dded5b74237…1e8c · Hugging Face

Download

tex_dec_f16.gguf

gguf · 905 MB · SHA-256 afd304f4dfcb…e4e0 · Hugging Face

Download

tex_slat_flow_1024_f16.gguf

gguf · 2.44 GB · SHA-256 bbb55b0910c7…58a8 · Hugging Face

Download

tex_slat_flow_512_f16.gguf

gguf · 2.44 GB · SHA-256 89a081b7f548…52ae · Hugging Face

Download

Source excerpts

3 excerpts
ReadmeAug 1huggingface.co/LocalAI-io/TRELLIS.2-4B-GGUF

--- license: mit base_model: microsoft/TRELLIS.2-4B tags: - gguf - trellis - image-to-3d - ggml pipeline_tag: image-to-3d --- # TRELLIS.2-4B — GGUF (f16) GGUF (`f16`) conversions of [microsoft/TRELLIS.2-4B](https://huggingface.co/microsoft/TRELLIS.2-4B) for the **[trellis2cpp](https://github.com/localai-org/trellis2cpp)** / [ggml](https://github.com/ggml-org/ggml) runtime — a CUDA-free, PyTorch-free C++ port of the TRELLIS.2 image-to-3D pipeline. Load these directly; no safetensors conversion or Python is needed at inference. ## Files | File | Pipeline stage | Source safetensors | |---|---|---| | `ss_flow_f16.gguf` | Sparse-structure flow (64³ occupancy) | `ss_flow_img_dit_1_3B_64_bf16` | | `slat_flow_f16.gguf` | Shape-SLAT flow, 512 fine | `slat_flow_img2shape_dit_1_3B_512_bf16` | | `slat_flow_1024_f16.gguf` | Shape-SLAT flow, 1024 cascade | `slat_flow_img2shape_dit_1_3B_1024_bf16` | | `shape_dec_f16.gguf` | Shape decoder → dual-grid fields | `shape_dec_next_dc_f16c32_fp16` | | `shape_enc_f16.gguf` | Shape encoder (re-encode for texture flow) | `shape_enc_next_dc_f16c32_fp16` | | `tex_dec_f16.gguf` | PBR texture decoder | `tex_dec_next_dc_f16c32_fp16` | | `tex_slat_flow_512_f16.gguf` | Texture-SLAT flow, 512 | `slat_flow_imgshape2tex_dit_1_3B_512_bf16` | | `tex_slat_flow_1024_f16.gguf` | Texture-SLAT flow, 1024 | `slat_flow_imgshape2tex_dit_1_3B_1024_bf16` | The image conditioning encoder (DINOv3) and the sparse-structure decoder are published separately: [dinov3-vitl16-pretrain-lvd1689m-GGUF](https://huggingface.co/LocalAI-io/dinov3-vitl16-pretrain-lvd1689m-GGUF) and [TRELLIS-image-large-GGUF](https://huggingface.co/LocalAI-io/TRELLIS-image-large-GGUF). ## Conversion Converted from the upstream `bf16`/`fp16` safetensors to GGUF `f16` (ggml ftype 1) with the `convert_*_to_gguf.py` scripts in the trellis2cpp project. No weights were retrained or modified — this is a format conversion only. ## License MIT, inherited from [microsoft/TRELLIS.2-4B](https://huggingface.co/microsoft/TRELLIS.2-4B). See [`LICENSE`](LICENSE). Copyright (c) Microsoft Corporation.

NoteOct 10huggingface.co/LocalAI-io/TRELLIS.2-4B-GGUF

Source context: 12198 downloads · 13 likes · Pipeline image-to-3d · Library trellis · Repo LocalAI-io/TRELLIS.2-4B-GGUF

NoteAug 1huggingface.co/LocalAI-io/TRELLIS.2-4B-GGUF

Source context: 1022 downloads · 3 likes · Pipeline image-to-3d · Library trellis · Repo LocalAI-io/TRELLIS.2-4B-GGUF