New technique gives users fine-grained control over AI image-editing strength

Researchers have developed Kontinuous Kontext, a new method for instruction-based image editing that adds a slider-like dimension to AI-powered photo manipulation. Rather than relying solely on text prompts, the model lets users specify how strong an edit should be—smoothly scaling from no visible change all the way to a fully realized transformation.

How it works#

The system extends an existing instruction-driven image editing model by accepting an additional input: a scalar value representing edit strength. A lightweight neural network (called a projector) translates this strength parameter and the text instruction into control signals that modulate the underlying model's behavior. This enables the model to apply edits at any intensity level without needing separate training for different strength levels.

Training and dataset#

To train the system, the researchers synthesized a dataset of image-instruction-strength combinations using existing generative models, then filtered the results to ensure quality and consistency. This synthetic approach allowed them to create diverse training examples across many editing scenarios.

Capabilities and applications#

Kontinuous Kontext handles a wide range of editing tasks—from stylization and attribute modifications to material, background, and shape changes—all within a single unified framework. The method works without requiring specialized, operation-specific training, making it flexible enough to adapt to different editing goals while maintaining user control over how pronounced each change becomes.