Hydrilla/BlueFox3D-retrained-characters

BlueFox3D retrained characters (4-stage L4 pack)

This is not a replacement for Hydrilla/BlueFox3D.
Default BlueFox3D inference still uses that stock repo.

This pack is four 1536-cascade denoisers for a small character-like Objaverse-LVIS subset:

Slot	Source	Retrained here?
SS-64	last-4-block finetune from stock SS-64	Yes (1000 steps)
Shape-512	copy of Hydrilla/BlueFox3D-character-shape512	No (not rerun)
Shape-1024	last-4-block finetune from stock shape-1024	Yes (1000 steps)
Tex-1024	last-4-block finetune from stock tex-1024	Yes (1000 steps)
SS / shape / tex VAEs	Hydrilla/BlueFox3D	No

This is not a full retrain of BlueFox3D. Same 149 meshes, last 4 of 30 blocks + out_layer, 8-bit Adam, 1000 steps per new stage. Do not call it a quality win until a hold-out compare says so.

What was trained

Item	Value
Trainable	last 4 transformer blocks + out_layer
Steps	1000 per new stage (50-step VRAM smoke first)
Batch	1 / GPU, max_tokens 2048 on 1024/tex
GPU	1× NVIDIA L4 24 GB (sequential; DDP does not pool VRAM)
Optimizer	AdamW 8-bit, lr 1e-5, no EMA
Shape-512	reused previous character finetune

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.

Split	Count	Used?
Train GLBs	150 downloaded, 149 dual-grid / SS / shape	Yes
PBR / tex	120 (shader dump failed on the rest)	Tex-1024 only
Hold-out GLBs	20	Never

Use

python inference.py --image assets/images/21_img.png --output ./pack.glb \
  --seed 42 --low_vram --resolution 1536 --preset baseline \
  --model_path Hydrilla/BlueFox3D-retrained-characters

--low_vram is required on 24 GB GPUs (~18–22 GB). Full 1536 without it needs ~40 GB+.

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 — SS-64 / shape-512 / shape-1024 / tex-1024 from this repo; decoders from Hydrilla/BlueFox3D
• ckpts/ss_flow_img_dit_1_3B_64_bf16.safetensors
• ckpts/slat_flow_img2shape_dit_1_3B_512_bf16.safetensors (copied, not retrained)
• ckpts/slat_flow_img2shape_dit_1_3B_1024_bf16.safetensors
• ckpts/slat_flow_imgshape2tex_dit_1_3B_1024_bf16.safetensors

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}
}