3DreamBooth-CustomBench Checkpoints

Pretrained joint (3DreamBooth + 3Dapter) LoRA checkpoints for all 30 subjects in 3D-CustomBench, released alongside 3DreamBooth so you can run subject-driven video generation without training from scratch.

Each subject was trained for 400 steps with train_joint.py. Folder names match the official 3D-CustomBench subject ids exactly.

Contents

subjects.json                                           # trigger phrase + metadata for all 30 subjects
<subject_id>/
├── subject_adapter/pytorch_lora_weights.safetensors     # 3DreamBooth identity LoRA
├── tdapter/pytorch_lora_weights.safetensors             # subject-specific 3Dapter LoRA
└── training_config.json                                 # training args for this subject

Trigger phrases

Every subject uses the rare identifier rhs followed by a class word that differs per subject — rhs bear, rhs mug, rhs bust, and so on. This phrase is what carries the learned identity, so it must appear verbatim in your inference prompt and be marked as the LoRA span. Using the wrong class word (e.g. rhs plushie for a checkpoint trained on rhs bear) will not activate the subject identity.

subjects.json is the machine-readable source for this:

import json
from huggingface_hub import hf_hub_download

meta = json.load(open(hf_hub_download(
    "lanikoworld/3DreamBooth-CustomBench", "subjects.json")))
spans = {s["subject_id"]: s["lora_span"] for s in meta["subjects"]}
print(spans["graduation_bear"])   # "rhs bear"

Each entry carries subject_id, lora_span, identifier, class_word, the exact train_prompt used, and the subject's benchmark_prompt from 3D-CustomBench.

Subjects and prompt spans

Subject id	LoRA span	Training prompt
bear_keychain	rhs plushie	A video of a [rhs plushie].
black_gold_mug	rhs mug	A video of a [rhs mug].
black_handbag	rhs handbag	A video of a [rhs handbag].
blue_label_pill_bottle	rhs bottle	A video of a [rhs bottle].
blue_pig_mug	rhs mug	A video of a [rhs mug].
cat_figurine	rhs figurine	A video of a [rhs figurine].
ceramic_bust	rhs bust	A video of a [rhs bust].
covered_motorcycle	rhs motorbike	A video of a [rhs motorbike].
deer_flowerpot	rhs pot	A video of a [rhs pot].
drawstring_pants	rhs sweatpants	A video of a [rhs sweatpants].
floral_mug	rhs mug	A video of a [rhs mug].
gaming_headset	rhs headset	A video of a [rhs headset].
graduation_bear	rhs bear	A video of a [rhs bear].
hand_cream	rhs tube	A video of a [rhs tube].
headband_bust	rhs bust	A video of a [rhs bust].
lavender_pitcher	rhs pitcher	A video of a [rhs pitcher].
lotion_bottle	rhs bottle	A video of a [rhs bottle].
milk_carton	rhs carton	A video of a [rhs carton].
moose_plush	rhs plush	A video of a [rhs plush].
multifunction_printer	rhs printer	A video of a [rhs printer].
office_chair	rhs chair	A video of a [rhs chair].
pink_plush	rhs plush	A video of a [rhs plush].
pink_rubber_duck	rhs duck	A video of a [rhs duck].
rattan_light_bulb	rhs bulb	A video of a [rhs bulb].
small_pill_bottle	rhs bottle	A video of a [rhs bottle].
textured_rock	rhs rock	A video of a [rhs rock].
toy_toilet	rhs toilet	A video of a [rhs toilet].
wafer_bag	rhs bag	A video of a [rhs bag].
white_light_bulb	rhs bulb	A video of a [rhs bulb].
yogurt_drink	rhs bottle	A video of a [rhs bottle].

subjects.json also records each subject's legacy_id — the internal name used during development — matching the legacy_id field in the dataset's manifest.json.

In every training_config.json, filesystem paths (pretrained_model_root, instance_data_root, reference_path, tdapter_path, output_dir) were rewritten from their original development locations to the public layout documented in the 3DreamBooth README. All hyperparameters and prompts are unmodified.

Usage

Download one subject:

hf download lanikoworld/3DreamBooth-CustomBench \
  --local-dir ./checkpoints/custombench --include "graduation_bear/*"

Run inference with validate_joint.py:

python validate_joint.py \
  --pretrained_model_root ./checkpoints/hunyuanvideo-1.5 \
  --pretrained_transformer_version 720p_t2v \
  --subject_adapter_path ./checkpoints/custombench/graduation_bear/subject_adapter/pytorch_lora_weights.safetensors \
  --tdapter_path ./checkpoints/custombench/graduation_bear/tdapter/pytorch_lora_weights.safetensors \
  --reference_path ./datasets/3d-custombench/subjects/graduation_bear/references \
  --prompt "A video of a rhs bear on a beach." \
  --text_lora_spans "rhs bear" \
  --video_length 81

See the main repo's Quick start section for the full setup (base model + 3Dapter checkpoint download, the config-driven scripts/run.py interface, and how the [rhs ...] bracket span format works).

License

Apache License 2.0, matching the main repository. Weights are derived from HunyuanVideo-1.5 fine-tuning and remain subject to the Tencent Hunyuan Community License.

Citation

@misc{ko20263dreambooth,
  title         = {3DreamBooth: High-Fidelity 3D Subject-Driven Video Generation Model},
  author        = {Hyun-kyu Ko and Jihyeon Park and Younghyun Kim and Dongheok Park and Eunbyung Park},
  year          = {2026},
  eprint        = {2603.18524},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2603.18524}
}