Hi3D v1.0 for ComfyUI
Hi3D v1.0 (SPARC3D) mesh refinement weights repackaged as single ComfyUI safetensors files.
| File | Folder | Contents |
|---|---|---|
checkpoints/hi3d_v1.safetensors |
models/checkpoints |
transformer_512, transformer_1024 and the mesh VAE in one checkpoint (fp16/fp32 as released) |
checkpoints/hi3d_v1_int8.safetensors |
models/checkpoints |
same, with the transformer block linears quantized to int8 (tensorwise, convrot) and int8 attention |
clip_vision/dinov2-large.safetensors |
models/clip_vision |
DINOv2-large image encoder |
clip_vision/clip-vit-large-patch14.safetensors |
models/clip_vision |
CLIP ViT-L/14 image encoder |
Install the ComfyUI-Hi3D node pack, load the checkpoint with Load Hi3D Checkpoint and the two encoders with Load CLIP Vision. The pack ships example workflows for a single refine pass and for the 512 β 1024 β 1536 chain.
Modifications
The Hi3D weights in checkpoints/ are modified from the official Hi3D v1.0 release
(https://github.com/hi3d-official/Hi3Dv1.0): the separate transformer_512, transformer_1024 and vae
weights were merged into one file with renamed keys, and in hi3d_v1_int8.safetensors the transformer
block linear layers are quantized to int8. Weight values are otherwise unchanged.
The CLIP Vision files are unmodified third-party models (DINOv2 by Meta, Apache 2.0; CLIP by OpenAI, MIT).
License
Hi3D is licensed under the Hi3D Open License Agreement Version 1.0. Copyright Β© 2026 OCEANO HK HOLDING LIMITED. All rights not expressly granted are reserved. The Hi3D trademarks are owned by OCEANO HK HOLDING LIMITED or its applicable affiliate(s). See LICENSE, NOTICE and THIRD_PARTY_LICENSES.md, including the Acceptable Use Policy in the LICENSE, which applies to every use of these weights.