aiPublished: 11 Sept 2026, 09:00Trust level: official

Researchers Introduce AcFlow to Control Diffusion Transformers

Researchers Introduce AcFlow to Control Diffusion Transformers

Summary

AcFlow offers fine-grained style control and concept suppression for text-to-image Diffusion Transformers during inference without modifying base weights.

Prepared by the newsroom with AI assistance; the original link is always provided

Researchers have introduced AcFlow, an inference-time controller designed to provide fine-grained control over text-to-image Diffusion Transformers (DiTs). The method guides intermediate layer image-token activations through a learned concept-conditioned velocity field while keeping the core DiT model frozen.

The framework addresses limitations of direct prompting, allowing users to adjust style intensity continuously and suppress unwanted visual concepts. In benchmark tests, AcFlow achieved a style-content alignment score of 0.5365/0.2860, outperforming the top baseline score of 0.4397/0.2684.

The authors have made the source code publicly available on GitHub.

Source: arXiv cs.CV

Read the original →

Latest news