BlueFox3D character shape-512 (L4 finetune)
This is not a replacement for Hydrilla/BlueFox3D.
Default BlueFox3D inference still uses that stock repo.
This repo contains a shape-512 denoiser finetuned on a small character-like Objaverse-LVIS subset. All other cascade stages (sparse structure, shape-1024, texture-1024) are still loaded from stock Hydrilla/BlueFox3D.
What was trained
| Item | Value |
|---|---|
| Stage | shape-512 only (ElasticSLatFlowModel, 30 blocks, 1.39B params) |
| Trainable | last 4 transformer blocks + out_layer (~183M / 1.39B) |
| Steps | 1000 |
| Batch | 1 mesh / step |
| GPU | 1× NVIDIA L4 24 GB |
| Optimizer | AdamW 8-bit, lr 1e-5, no EMA |
| Start weights | public Pixal3D / BlueFox3D shape-512 |
1536 inference still runs:
- SS-64 — stock
- Shape-512 — this file
- Shape-1024 — stock
- Texture-1024 — stock
Expect a small mid-stage geometry shift, not a new product model.
Dataset
Objaverse-LVIS only. We did not download the ~8.9 TB allenai/objaverse snapshot.
LVIS keys (no class named character): person, teddy bear, doll, figurine, statue, mannequin, puppet, action figure, toy soldier, snowman, scarecrow. Matching was a bit loose, so extra toy/bear IDs mixed in.
| Split | Count | Trained on? |
|---|---|---|
| Train GLBs | 150 downloaded, 149 encoded (1 mesh failed voxelize) | Yes |
| Hold-out GLBs | 20 | Never |
| Encoded | 2 Blender views + dual-grid 512 + shape-512 view latents | Yes |
| PBR / SS latents | not built | No |
One-image compare (honest)
Same photo assets/images/21_img.png (jester), seed 42, --preset baseline, --resolution 1536, --low_vram.
| Stock Pixal3D / BlueFox3D | This shape-512 | |
|---|---|---|
| Weights | Hydrilla/BlueFox3D | this repo for shape-512 only |
| Vertices | 210,075 | 198,274 |
| Faces | 246,515 | 235,199 |
The two GLBs are not identical, but they look similar. This is not a quality-win claim. 21_img is a demo photo, not one of the 20 hold-out meshes.
Use
python inference.py --image assets/images/21_img.png --output ./ft.glb \
--seed 42 --low_vram --resolution 1536 --preset baseline \
--model_path Hydrilla/BlueFox3D-character-shape512
--low_vram is required on 24 GB GPUs.
Stock (unchanged):
python inference.py --image assets/images/21_img.png --output ./stock.glb \
--seed 42 --low_vram --resolution 1536 --preset baseline \
--model_path Hydrilla/BlueFox3D
Files
pipeline.json— shape-512 from this repo; other stages fromHydrilla/BlueFox3Dckpts/slat_flow_img2shape_dit_1_3B_512_bf16.safetensors— converteddenoiser_step0001000.ptckpts/slat_flow_img2shape_dit_1_3B_512_bf16.json— same architecture as stock shape-512
Code
Training and convert scripts live in HydrillaAI/Bluefox3D.
License / citation
MIT, same as BlueFox3D. Upstream: Pixal3D / TRELLIS.2.
@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}
}