Instructions to use t8star/Hi3d-Comfy with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use t8star/Hi3d-Comfy with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("t8star/Hi3d-Comfy", dtype=torch.bfloat16, device_map="cuda") prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k" image = pipe(prompt).images[0] - Notebooks
- Google Colab
- Kaggle
Hi3D T8 â Core and Workbench Models
English | įŽäŊ䏿
By T8star. Verified redistribution of the shared Hi3D core, geometry and PBR model components used by the standalone Hi3D T8 workbench. Original weights, configurations, pinned revisions and individual licenses are retained. UltraTex uses an upstream-published fp8 base; this project did not quantize it.
ComfyUI only uses the Hi3D core for reference-image + coarse-mesh geometry refinement. Image-to-3D and PBR extensions belong to the standalone workbench. This model repository contains neither the application nor Python. GitHub Releases contain code only; T8star will provide the complete package with Python and models separately.
Native nodes infer directly inside ComfyUI's own Python, without a workbench service. Upload the reference and coarse mesh to refine with the same original core weights. Replace legacy bridge canvases with the new example.
Workbench v0.6.0 adds Idol and asset-studio models. Existing core weights are retained. New features require the v0.6 runtime; a model download or code ZIP alone does not upgrade an older Python bundle. The optional Windows CUDA converter in runtime-wheels/v0.4.9/ is separately installed code, not a checkpoint or Python distribution.
Included models
| Directories | Purpose | License |
|---|---|---|
hi3d/ |
Hi3D V1.0 geometry refinement; 512 and 1024 checkpoints, shared VAE; 1536 uses 1024 weights | Hi3D Open License v1.0 |
dino/, clip/ |
Required DINOv2 Large and CLIP ViT-L/14 core encoders | Apache-2.0 / MIT |
extensions/coarse/ |
Hunyuan3D-2mini Turbo image-to-geometry | Tencent Hunyuan 3D 2.0 Community |
extensions/coarse_hunyuan21/ |
Hunyuan3D 2.1 image-to-geometry | Tencent Hunyuan 3D 2.1 Community |
extensions/coarse_hunyuan_mv/ |
Hunyuan3D-2mv multi-view geometry: front with left/back/right views | Tencent Hunyuan 3D 2.0 Community |
extensions/coarse_idol/, extensions/idol_background/ |
Idol diffusion/decoder, DINO-H+, FLUX.2 VAE and BiRefNet | Apache-2.0 / DINOv3 License / MIT |
extensions/coarse_trellis2/, extensions/trellis_structure/ |
TRELLIS.2 geometry and structure decoder | MIT |
extensions/texture/, extensions/texture_encoder/ |
TRELLIS.2 PBR and DINOv3 encoder | MIT / DINOv3 License |
extensions/texture_hunyuan21/, extensions/paint_dino/, extensions/paint_upscaler/ |
Hunyuan3D-Paint 2.1, DINOv2 Giant and Real-ESRGAN | Tencent / Apache-2.0 / BSD-3-Clause |
extensions/texture_ultratex/, extensions/ultratex_backbone/, extensions/ultratex_text_encoder/ |
UltraTex adapters/decoder, FLUX.2 Klein 9B fp8, Qwen3 and VAE | MIT / FLUX Non-Commercial / Apache-2.0 |
extensions/ultratex_backbone_int8/ |
tsolful Klein Base 9B INT8 ConvRot, pinned CivitAI version 3080042/file 2959286 | Original FLUX Non-Commercial |
extensions/background/ |
Optional CPU U2Netp foreground extraction | MIT / Apache-2.0 |
studio/p3sam/, studio/parts_full/, studio/sonata/ |
Part segmentation and X-Part completion | Tencent Hunyuan 3D-Part Community / Apache-2.0 |
studio/skintokens/, studio/qwen_rig_config/, studio/unirig/ |
Character skeleton and skin prediction | MIT / Apache-2.0 |
studio/kimodo/, studio/text_base/, studio/text_adapter/, studio/text_adapter_supervised/ |
Single-character text motion and LLM2Vec encoders | NVIDIA Open Model / Llama 3 Community |
manifest.json, extensions/manifest.json and studio/manifest.json record 3 core, 17 geometry/PBR and 10 asset-studio components. model-catalogue.json lists all 169 model/config/license artifacts (122.89GB) with byte counts and SHA256. ComfyUI and the workbench use identical Hi3D weights; no separate ComfyUI checkpoint exists. The INT8 file is the community-published Base quantization, not distilled/KV; it was not quantized by this project. FP8 remains available separately.
2026-11-08 update: Idol adds 7 artifacts (3.78GB); asset-studio adds 63 artifacts (34.91GB). Multi-view modeling remains available. Select Idol for native Gaussian outputs and optional colored mesh/PBR, or choose Hunyuan3D-2mv for front/left/back/right references. Subsequent refinement and existing painters use the front reference.
New controls require v0.6.0 or newer with its matching complete runtime. Release and code. GitHub assets contain no Python or weights; the author distributes the full runtime bundle separately.
Studio models are now hosted under studio/, retaining their pinned official repositories and licenses. Built with Meta Llama 3: Kimodo's LLM2Vec text encoder uses the unchanged Meta-Llama-3-8B-Instruct base and two adapters. Retain studio/text_base/LICENSE, USE_POLICY.md and the component notices. Runtime wheels are counted separately from the 169 model artifacts.
Download
For native ComfyUI, use ComfyUI/models/Hi3D/. For the workbench, use its models/ directory. The same core weights can also be selected through a custom model root. Replace models below with the desired target.
Core only (10.66GB), sufficient for the independent Hi3D ComfyUI nodes:
hf download t8star/Hi3d-Comfy --local-dir models --include "manifest.json" "hi3d/*" "dino/*" "clip/*" "licenses/*"
All hosted core, geometry/PBR and studio files (about 122.89GB; allow additional runtime and working space):
hf download t8star/Hi3d-Comfy --local-dir models --exclude "runtime-wheels/*"
Download the complete component set in one row for each workbench engine you need:
| Workbench selection | Directories under extensions/ |
Download size |
|---|---|---|
| Hunyuan3D-2mini Turbo geometry | coarse |
3.82GB |
| Hunyuan3D 2.1 geometry | coarse_hunyuan21 |
7.37GB |
| Hunyuan3D-2mv multi-view geometry | coarse_hunyuan_mv |
4.93GB |
| Idol Gaussian/colored mesh generation | coarse_idol, idol_background |
3.78GB |
| TRELLIS.2 geometry | coarse_trellis2, trellis_structure, texture_encoder |
10.06GB |
| TRELLIS.2 PBR | texture, texture_encoder |
8.04GB |
| Hunyuan3D-Paint 2.1 | texture_hunyuan21, paint_dino, paint_upscaler |
11.50GB |
| UltraTex FP8 | texture_ultratex, ultratex_backbone, ultratex_text_encoder |
19.62GB |
| UltraTex INT8 ConvRot | texture_ultratex, ultratex_backbone_int8, ultratex_text_encoder |
19.46GB |
Sizes use decimal GB, include each row's shared dependencies, and exclude Python, temporary files and optional accelerators. Shared directories download once, so do not add all rows to estimate the full repository. Hi3D refinement is optional in the workbench; add the core only if you enable it. Background removal adds extensions/background/ (about 4.6MB).
| Asset-studio operation | Directories under studio/ |
Download size |
|---|---|---|
| P3-SAM segmentation | p3sam, sonata |
0.89GB |
| X-Part completion | parts_full, p3sam, sonata |
9.93GB |
| SkinTokens rigging | skintokens, qwen_rig_config |
1.62GB |
| UniRig rigging | unirig |
5.82GB |
| Kimodo text motion | kimodo, text_base, text_adapter, text_adapter_supervised |
17.55GB |
For example, download Idol and its own foreground model without the other engines:
hf download t8star/Hi3d-Comfy --local-dir models --include "extensions/manifest.json" "extensions/coarse_idol/*" "extensions/idol_background/*" "licenses/idol/*"
Studio files belong under models/studio/; include studio/manifest.json and every component in the relevant row. Mesh simplification and motion retargeting do not require new neural weights. Operation and quality limits.
For example, TRELLIS.2 geometry and PBR, with their shared encoder:
hf download t8star/Hi3d-Comfy --local-dir models --include "extensions/manifest.json" "extensions/coarse_trellis2/*" "extensions/trellis_structure/*" "extensions/texture/*" "extensions/texture_encoder/*" "licenses/*"
The workbench's Models page can also download only the required engine components from pinned original/public sources. Missing optional engines do not disable installed ones. Verify files in the workbench before inference.
To use another drive or shared location, select its root in Models and environment â Model directory and use that same path with hf download --local-dir. Keep hi3d/, dino/, clip/ and extensions/ directly under the selected root. The idle switch takes effect without restarting; it does not move existing files. Relative paths resolve against the workbench root, and code updates preserve the saved setting.
Existing core files at a newly selected location still require Reverify all core models before that location is reported ready.
Program updates
The standalone bundle checks stable GitHub Releases, downloads code updates and installs them on the next launch. Models, Python runtimes, configuration and history are preserved. Code updates do not download this entire model repository. Runtime-incompatible versions require a new complete package. Update instructions.
Acceleration packages are code, not model weights. The optional, hash-pinned hi3d-voxel-cuda wheel is stored under runtime-wheels/v0.4.9/; the workbench installs the compatible build separately. FlashAttention and Numba/llvmlite are obtained from the external sources pinned by the installer. SDPA and the official CPU converter remain available. Models never include a bundled Python environment. See the performance guide.
Validation and limits
Real model runs were tested on Windows / RTX 5090 Laptop 24GB / 64GB RAM. Smaller GPUs have not been validated. Hi3D refinement is optional, rebuilds topology and does not generate textures; the separate PBR stage exports embedded-material GLB. Geometry, back views, UV seams and materials require inspection; printability and universal quality improvements are not guaranteed. Engine evidence.
Licenses
Each component retains its own license and attribution; no single MIT license covers the collection. Hi3D's LICENSE/NOTICE/AUTHORS are retained. DINOv3 retains the Meta agreement. UltraTex's FLUX.2 Klein 9B base uses the FLUX Non-Commercial License. See component directories, licenses/ and the source manifests. Powered by Hi3D 1.0.
T8star
Bilibili ¡ YouTube ¡ API ¡ Free Gallery ¡ Online AI Apps ¡ ComfyUI Distribution ¡ Hugging Face
- Downloads last month
- -