OneActor++: Consistent Multi-Subject Generation via Cluster-Conditioned Guidance

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Text-to-image diffusion models offer high-quality image generation but struggle with subject consistency.
• Previous methods like OneActor efficiently generate single consistent subjects but lack multi-subject capabilities.
• Existing models cannot synthesize multiple consistent subjects from user-provided image references.

Purpose of the Study:

• Introduce OneActor++, a novel paradigm for consistent multi-subject image generation.
• Enable the synthesis of multiple subjects using both textual prompts and image references.
• Address the demand for generating multiple consistent subjects beyond the scope of prior work.

Main Methods:

• Pioneer a multi-target formulation of cluster guidance theory for multi-subject identity preservation.
• Employ a Mixture of Projectors (MoP) architecture for holistic multi-subject representation.
• Utilize an enhanced tuning recipe with anchored noising and drift regulation.
• Incorporate a one-step inversion mechanism for image-referenced generation.

Main Results:

• OneActor++ significantly outperforms existing baselines in generating multiple consistent subjects.
• Achieves excellent subject consistency and superior prompt alignment.
• Demonstrates high image quality in generated outputs.
• Successfully enables multi-subject generation from both text and image inputs.

Conclusions:

• OneActor++ is a powerful paradigm for consistent multi-subject image generation.
• Offers significant advancements over previous methods for AI-driven art creation.
• Shows strong practical potential across diverse application scenarios.