HOMIE¶

Development line: project:homie · thread homie-public-project-release
Last event: 2026-07-22 · 1 dated since 2026-07-22 · Researched: 2026-09-05 · confidence: medium

What it is¶

HOMIE is an open-source local video generation pipeline driven by reference images of people, products, logos, OCR maps, and multi-view shots.

• Multi-subject composition: combines multiple distinct subjects in one video clip.
• Multi-reference alignment: uses several references of one subject for text accuracy or viewpoint consistency.

Running it requires Wan2.1-T2V-14B, separate HOMIE weights at 37.1 GB, and Qwen3-VL-2B-Thinking.

This is a research inference release, not a cloud service.

Development line¶

• 2026-07-22 — HOMIE public project resources appeared. On 2026-07-22, the project linked its website, GitHub repository, and Hugging Face model page homie-r2v-wan2.1 together. These links provide public documentation, source code, and model distribution in one place. Precise release claims, version details, capabilities, and license are not stated.

What changed¶

2026-07-20 — The HOMIE technical report appeared on arXiv, establishing a unified method for inter- and intra-subject video personalization.

2026-07-21 — Inference code and checkpoints fine-tuned on Wan2.1-T2V-14B were published.

2026-07-22 — The project page, code, and weights described a single local workflow, with no separate version change found on this day.

How to use this¶

From 2026-07-22, start with the linked project page, source repository, and Hugging Face model page. Verify the project documentation, requirements, and license before use.

1. Install the environment using set_env.sh. This requires PyTorch 2.4.0 or newer with a compatible CUDA build. — https://github.com/YIYANGCAI/HOMIE
2. Download Wan-AI/Wan2.1-T2V-14B-Diffusers, yychai/homie-r2v-wan2.1 weights, and Qwen/Qwen3-VL-2B-Thinking into the specified directories. — https://github.com/YIYANGCAI/HOMIE
3. Create a JSONL file: outer lists in reference_paths separate subjects, and prompt describes the video. — https://github.com/YIYANGCAI/HOMIE
4. Run generate_mllm_feature.py, then pass the generated JSONL with mllm_feature paths to generate.py. — https://github.com/YIYANGCAI/HOMIE

Best practices¶

• Do not skip feature extraction with Qwen3-VL-2B-Thinking: inference expects a JSONL with added mllm_feature paths. — https://github.com/YIYANGCAI/HOMIE
• Start with the documented single-GPU mode: 832×480, 97 frames, 24 fps, and a fixed base_seed. Apply FSDP and context parallel only on a suitable multi-GPU node. — https://github.com/YIYANGCAI/HOMIE
• For context parallel, use the documented 8×A100 example with ULYSSES_SIZE=8 and RING_SIZE=1. This is a sample configuration, not a stated hardware minimum. — https://github.com/YIYANGCAI/HOMIE

Superseded by this¶

• Nothing marked obsolete yet.

Still unknown¶

• The repository lacks versioned releases or a changelog: the exact commit or weight corresponding to the 2026-07-22 entry is not established.
• Official minimum VRAM, generation time, and tested configurations beyond the 8×A100 example are not published.
• The current README includes a context-parallel scenario, but the source does not date its addition, so it is not listed as a separate dated event.

Sources¶

Agent brief¶

• Subject: project:homie, thread homie-public-project-release, 1 dated events 2026-07-22 → 2026-07-22.
• Practical note: From 2026-07-22, practitioners evaluating HOMIE should begin with the linked project page, source repository, and Hugging Face model page, then validate the project documentation, requirements, and license before use.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.