arXiv:2510.09060v2 Announce Type: replace Abstract: Flow-based text-to-image models follow deterministic trajectories, making it costly to explore diverse modes under limited sampling budgets. Existing approaches to improving diversity often rely on retraining or degrade image fidelity. To address this limitation, we present a training-free, inference-time control mechanism that makes the flow itself diversity-aware. Our core insight is to encourage diversity through guidance that is geometrically decoupled from the mode's quality-seeking direction. Our method simultaneously encourages lateral
The continuous evolution of generative AI models necessitates novel approaches to address limitations like diversity in outputs, especially as these models become more complex and widespread.
This development allows for improved control over the diversity of outputs from text-to-image models without compromising fidelity, addressing a key challenge in generative AI applications.
Generative AI models can now produce a wider range of diverse outputs more efficiently and without requiring costly retraining or sacrificing quality, making them more versatile for creative and exploratory tasks.
- · Generative AI developers
- · Digital content creators
- · AI research institutions
- · Creative industries
- · N/A
Improved diversity in AI-generated imagery leads to more innovative and varied creative applications.
The ability to produce diverse outputs efficiently could accelerate the development of more nuanced and specialized AI models for various artistic and design fields.
Enhanced control over generative AI outputs may lead to new forms of intellectual property and challenges in attributing creativity within AI-assisted works.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.AI