Homie generates consistent videos from multi-subject and object references

Homie is an open-source framework for human-object centric video personalization. Basically, it takes multiple reference images of specific people and items and turns them into consistent, high-fidelity videos without warping key details.

What it does

The model handles multiple subjects in a single generation, supporting both photorealistic human references and stylized 3D animation characters.

It excels with complex designs, like detailed patterns on a kimono or intricate toys, where traditional video generators usually hallucinate or lose consistency.

For creator marketing, you can feed it a single photo of an influencer alongside a product image to batch-generate realistic UGC-style promotional videos.

Its standout feature is intra-subject reference support. You can upload multiple viewing angles of the same character or object, giving the model enough spatial information to keep appearances stable during camera pans and turns.

Homie also accepts OCR layout maps. By telling the model exactly where text or logos should sit on an item, it dramatically improves text fidelity on generated products.

How it works

Personalized video generation typically struggles when multiple subjects interact, often resulting in awkward copy-paste artifacts or mangled logos.

Previous methods either crammed multimodal features into a constrained text embedding space, bottlenecking visual reasoning, or replaced the text encoder entirely with an MLLM, incurring massive re-alignment overhead.

Homie keeps the text encoder intact to preserve text prompt controllability, allowing its integrated MLLM to focus purely on understanding reference relationships.

It injects global multimodal guidance into self-attention layers to align MLLM semantic features with VAE tokens, using modality-reference embeddings to keep tokens organized.

Compared to earlier personalization models like VACE and Phantom, Homie shows noticeably higher fidelity and fewer visual errors, particularly on complex multi-subject scenes.

The team has released both the code and model weights publicly for developers to test.

Why it matters

Homie tackles two of the nastiest friction points in AI video generation: keeping complex items consistent across angles and placing brand logos accurately. For commercial marketers and e-commerce teams wanting scalable UGC-style ads, this moves generative video closer to practical production pipelines.

Its hybrid architecture is also a smart design lesson for the industry, showing that delegating reference reasoning to an MLLM while keeping classic text encoders beats an expensive, scorched-earth model overhaul. The main hurdle now is pipeline preparation, since users need multi-view images or OCR maps to unlock its highest fidelity.