3DreamBooth¶

Development line: project:3dreambooth · thread 3dreambooth-development
Last event: 2026-03-21 · 1 dated since 2026-03-21 · Researched: 2026-09-07 · confidence: medium

What it is¶

3DreamBooth is a research codebase for adapting HunyuanVideo-1.5 to a subject captured from multiple viewpoints.

• 3DreamBooth: one-frame, per-subject LoRA optimization for spatial identity.
• 3Dapter: a reference-image conditioning adapter jointly optimized with the subject LoRA.

The paper reports 720p, 81-frame, 50-step outputs; the public repository still marks evaluation code unavailable. Use it for controlled reproduction or supplied-checkpoint inference, not as a turnkey production package.

Development line¶

• 2026-03-21 — 3DreamBooth project resources were publicly linked. On 2026-03-21, an entry linked the 3DreamBooth project website and its GitHub source repository. We treat this as a public reference point for project history. We do not infer an unevidenced release, feature, or other change.

What changed¶

• 2026-03-19 — arXiv v1 introduced 3DreamBooth, 3Dapter, and 3D-CustomBench.
• 2026-03-20 — The repository records its initial commit, removes the Code Coming Soon badge, and uploads source files.
• 2026-03-21 — The project page and repository were linked for the dated event. We have no independent verification of an official launch date.
• 2026-07-22 — Hugging Face listed the 3D-CustomBench dataset as updated.
• 2026-07-28 — Hugging Face listed the 3DreamBooth-CustomBench checkpoint collection as updated.

How to use this¶

1. Clone the repository, create its Python 3.10 environment, and install the declared requirements. — https://github.com/Ko-Lani/3DreamBooth
2. Download HunyuanVideo-1.5, its text and vision dependencies, and the pretrained 3Dapter. Obtain access to gated FLUX.1-Redux-dev before launching a job. — https://github.com/Ko-Lani/3DreamBooth
3. For a ready-made trial, download one of the 30 joint CustomBench checkpoint pairs and use its exact identifier-plus-class trigger phrase. — https://huggingface.co/lanikoworld/3DreamBooth-CustomBench
4. For a custom subject, prepare ordered multi-view captures in images/ and a smaller set of background-removed, white-normalized square references in references/. — https://github.com/Ko-Lani/3DreamBooth
5. Run the supplied YAML configuration with --dry-run first, then use the 3dreambooth baseline or joint mode as appropriate. — https://github.com/Ko-Lani/3DreamBooth/blob/main/docs/training-modes.md
6. Keep generated video length at 4n + 1, such as 49, 81, or 129 frames. Do not treat a visual run as official metric reproduction while evaluation code is absent. — https://github.com/Ko-Lani/3DreamBooth

Best practices¶

• Start with a supplied CustomBench checkpoint before attempting custom training. It isolates environment and prompt errors from subject-training errors. — https://huggingface.co/lanikoworld/3DreamBooth-CustomBench
• Use diverse, ordered views and references covering the full 360-degree extent of the object. Remove reference backgrounds before conditioning. — https://arxiv.org/html/2603.18524
• Keep the bracketed identifier-plus-class phrase identical across training and validation. The literal phrase is the LoRA span that activates identity. — https://huggingface.co/lanikoworld/3DreamBooth-CustomBench
• Validate YAML resolution with --dry-run before loading large checkpoints or starting distributed training. — https://github.com/Ko-Lani/3DreamBooth/blob/main/docs/training-modes.md
• Check FLUX access and Hunyuan-derived licensing terms before committing to a workflow or commercial use. — https://github.com/Ko-Lani/3DreamBooth

Superseded by this¶

• 2026-03-20 — The initial Code Coming Soon state is obsolete: repository history records its removal and source-file uploads.

Still unknown¶

• We found no immutable tag or GitHub Release for the exact code state on 2026-03-21, so we cannot pin that revision from inspected public pages.
• The July Hugging Face labels establish update dates, but not the exact file-level changes made on those dates.
• We performed no local installation, GPU-memory measurement, throughput test, or independent metric reproduction; official evaluation code remains unavailable.
• We found no first-party Simplified Chinese documentation or official Chinese community guidance; current practical guidance rests on English primary sources.

Sources¶

Agent brief¶

• Subject: project:3dreambooth, thread 3dreambooth-development, 1 dated events 2026-03-21 → 2026-03-21.
• Practical note: See the sourced usage and practice sections above, including their limits.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.