SceneCraft: Interactive System for Image Editing via Scene Graph

Researchers have introduced SceneCraft, an interactive framework that lets users edit images by manipulating a visual scene graph instead of writing text prompts, according to a paper posted to arXiv on June 15, 2026 [1][2]. The system represents an image as an editable graph of objects and their relationships. Users perform spatial and relational operations directly on the graph, and those modifications are automatically translated into context-aware editing prompts [1][2]. The structured prompts are then dispatched to multiple generative models to produce the final output [1][2]. The authors argue this approach eliminates the linguistic ambiguity that often frustrates natural-language image editing, particularly in scenes with several interacting objects [2]. Evaluations showed that outputs generated through SceneCraft were consistently rated higher in quality and fidelity than those produced through manual prompt engineering [1][2]. Generative artificial intelligence has advanced rapidly in recent years, with models now capable of producing text, images, and video [6]. However, text-to-image editing tools have typically required users to iterate through trial-and-error prompting to achieve desired results [2]. SceneCraft shifts the interaction from linguistic guesswork to direct manipulation of a structured visual representation [1][2]. The paper arrives amid broader scrutiny of AI-generated and AI-edited imagery. Deepfakes—media edited or generated using artificial intelligence—have drawn attention for their potential use in disinformation, fraud, and non-consensual pornography [7]. While SceneCraft is presented as a creative tool, any system that lowers the barrier to realistic image editing also intersects with ongoing debates about synthetic media detection and provenance [7]. The work was posted on arXiv, the open-access e-print repository operated by Cornell University that provides rapid dissemination of research findings at no cost to readers or submitters [9][11]. The paper appears under the Computer Vision and Pattern Recognition category [1].