Hi3D V1.0 — T8 Model Bundle
English | 简体中文
By T8star. Verified, unmodified Hi3D V1.0 weights and required image encoders for ComfyUI-Hi3D T8 and the local workbench. This is a community redistribution, not a new model, fine-tune or quantization.
Hi3D refines a reference image and a coarse mesh of the same object. It does not generate the initial coarse mesh, UVs or textures. No other 3D generation models are included.
Files and provenance
| Directory | Original model | Pinned revision | License |
|---|---|---|---|
hi3d/ |
hi3d-official/Hi3Dv1.0 | 705fc0901c5d896cdcbf1dbf1feaa173e62bdd3a |
Hi3D Open License v1.0 |
dino/ |
facebook/dinov2-large | 47b73eefe95e8d44ec3623f8890bd894b6ea2d6c |
Apache-2.0 |
clip/ |
openai/clip-vit-large-patch14 | 32bd64288804d66eefd0ccbe215aa642df71cc41 |
MIT |
manifest.json records the SHA256 and byte size of every distributed model artifact. The model/configuration files are identical to the offline workbench. DINOv2 and CLIP are required image encoders, not additional 3D models. transformer_512 serves 512 refinement; transformer_1024 serves both 1024 and 1536. There is no separate 1536 checkpoint.
Download and use
Run from the workbench root:
hf download t8star/Hi3d-Comfy --local-dir models
This produces models/hi3d/, models/dino/, models/clip/ and models/manifest.json. Start the configured workbench backend and wait for SHA256 verification. In ComfyUI, load the canvas workflow. Weights do not belong in ComfyUI/models for this integration.
ComfyUI and the standalone workbench use the same weights and inference backend. The Registry package contains node code and the canvas workflow; the complete offline workbench also contains Python, dependencies and models. This repository contains models and licenses only, not a standalone application. Setup: English / 中文.
Validation and limits
Real 512 ComfyUI and 512/1024/1536 workbench runs were validated on Windows / RTX 5090 Laptop 24GB / 64GB RAM. Smaller GPUs have not been validated. Start with 512, 25 steps and CPU offload. Windows uses an explicit SDPA compatibility adapter; the official Linux path uses FlashAttention. See test evidence and limitations.
Outputs are geometry only. Hi3D normalizes scale and rebuilds topology; watertightness, printability and quality improvements are not guaranteed.
License
Hi3D files retain the Hi3D Open License v1.0, including its attribution, acceptable-use and scale conditions. Original hi3d/LICENSE, hi3d/NOTICE, hi3d/AUTHORS and hi3d/THIRD_PARTY_LICENSES.md are preserved. Encoder licenses are in licenses/. No MIT license is asserted for the bundle as a whole. Powered by Hi3D 1.0.
T8star
Bilibili · YouTube · API · Free Gallery · Online AI Apps · ComfyUI Distribution · Hugging Face