BlueFox3D 1.0 (Hydrilla/BlueFox3D)

BlueFox3D generates textured 3D meshes with PBR maps (base color, metallic, roughness, alpha) from a single image or posed multi-view images, lifting pixel features into 3D through back-projection.

  • Code & training scripts: HydrillaAI/Bluefox3D
  • Model card (Hydrilla): hydrilla.ai/research#model-card
  • Weights provenance: Baseline Pixal3D / TRELLIS.2-architecture 1.3B 3-stage cascade weights hosted under the Hydrilla organization as the default checkpoint for BlueFox3D.

Optional experimental character L4 finetune pack (not a drop-in quality replacement):


Architecture

3-stage flow-matching cascade (1536_cascade default):

  1. Sparse Structure (64): ckpts/ss_flow_img_dit_1_3B_64_bf16 + ckpts/ss_dec_conv3d_16l8_fp16
  2. Shape SLat (512 β†’ 1024): ckpts/slat_flow_img2shape_dit_1_3B_{512,1024}_bf16 + ckpts/shape_dec_next_dc_f16c32_fp16
  3. Texture SLat (1024): ckpts/slat_flow_imgshape2tex_dit_1_3B_1024_bf16 + ckpts/tex_dec_next_dc_f16c32_fp16

Image conditioning uses DINOv3 (facebook/dinov3-vitl16-pretrain-lvd1689m) with MoGe-2 camera estimation and BiRefNet background removal.


Quickstart

Install TRELLIS.2 CUDA extensions (FlexGEMM, o_voxel, flash-attn, nvdiffrast) first, then clone HydrillaAI/Bluefox3D:

git clone https://github.com/HydrillaAI/Bluefox3D.git BlueFox3D1.0
cd BlueFox3D1.0
pip install -r requirements.txt
pip install "huggingface_hub>=0.34,<1.0"

Single-image inference (--low_vram is required on 24 GB GPUs such as NVIDIA L4 / RTX 4090):

python inference.py --image assets/images/1_img.png --output ./output.glb --low_vram --resolution 1536

Multi-view inference (transforms.json in Blender/NeRF camera-to-world format):

python inference_mv.py --views_dir assets/mv_images/example --output ./output_mv.glb --low_vram

Hardware Notes

  • 24 GB GPUs (NVIDIA L4 / RTX 3090 / RTX 4090): Supported at 1536 with --low_vram (stages offloaded to CPU between steps; ~4.2 min warm sequential average per asset, ~6–9 min cold CLI run).
  • 40 GB+ GPUs (A100 / H100): Can run without --low_vram for faster execution.

Benchmarks & Comparisons

15-Image Open-Weights Benchmark (Single NVIDIA L4, 24 GB VRAM, Seed 42)

Measured directly from generated GLBs and run metadata across 15 diverse images (objects, characters, scenes/dioramas):

Model Checkpoint / Source Completed Avg Time / Asset Avg Vertices Avg Faces Avg GLB Size Output Type
BlueFox3D (Stock) Hydrilla/BlueFox3D 15 / 15 4.2 min (254 s) 199,836 214,576 11.3 MB 2048 PBR maps
BlueFox3D (Char Pack) Hydrilla/BlueFox3D-retrained-characters 15 / 15 4.4 min (266 s) 203,623 214,524 11.7 MB 2048 PBR maps
TRELLIS.2-4B microsoft/TRELLIS.2-4B 15 / 15 9.4 min (564 s) 219,553 239,819 19.7 MB 2048 PBR maps
Hunyuan3D-2.1 tencent/Hunyuan3D-2.1 15 / 15 11.8 min (709 s) 31,939 40,000 6.2 MB Painted mesh (40k remesh)
Hi3DGen Stable-X/trellis-normal-v0-1 15 / 15 1.5 min (88 s) 426,673 853,476 19.5 MB Clay mesh (no PBR)

Observed trade-offs & methodology (no uncomputed Chamfer/CLIP/FID scores):

  • BlueFox3D (Stock): Recommended default. Preserves input color fidelity with 2048 PBR maps and completes all 15/15 on a 24 GB L4 at 1536 (--low_vram); observed artifacts include a stray staff on 21_img, a base slab on 5_img, and rug fragments on s_18_img.
  • BlueFox3D (Char Pack): Very close to stock on most images, reproduces stock artifacts on 21_img / 5_img, and fragmented worse on the isolated 21_img check (experimental only).
  • TRELLIS.2-4B: Strong geometry and 2048 PBR maps, with larger GLBs (19.7 MB avg) and 2.2Γ— slower runtime (9.4 min avg) on L4.
  • Hunyuan3D-2.1: Lightest GLBs (40k faces, 6.2 MB avg); averaged 11.8 min in this 5-model batch (3.9 min in an earlier warm 4-model run); failed on s_15_img (invented table disc), lost objects to background removal on s_18_img, and shifted colors on 7_img / 9_img.
  • Hi3DGen: Fastest runtime and sharpest clay geometry (853,476 avg faces), with no PBR textures (requires a separate texturing step).

BlueFox3D Variants Compared (21_img.png, Seed 42, 1536 --low_vram)

Repo What Changed Vertices Faces GLB Size Status
Hydrilla/BlueFox3D Stock 1.3B 3-stage cascade 211,389 246,891 14.0 MB Recommended default
Hydrilla/BlueFox3D-retrained-characters 4-stage L4 pack (SS-64, shape-512, shape-1024, tex-1024) 197,039 236,397 13.2 MB Experimental

License & Attribution

Released under the MIT License. Built on Pixal3D and TRELLIS.2; see NOTICE for third-party attribution.

@article{li2026pixal3d,
    title={Pixal3D: Pixel-Aligned 3D Generation from Images},
    author={Li, Dong-Yang and Zhao, Wang and Chen, Yuxin and Hu, Wenbo and Guo, Meng-Hao and Zhang, Fang-Lue and Shan, Ying and Hu, Shi-Min},
    journal={arXiv preprint arXiv:2605.10922},
    year={2026}
}
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for Hydrilla/BlueFox3D

Finetunes
1 model

Paper for Hydrilla/BlueFox3D