zoezheng126/slideforge-sam3-decoder

SlideForge SAM3 slide-component decoder

Fine-tuned SAM3 decoder for perception-aligned slide decomposition (30.4M trainable parameters, 3.6% of the model; frozen vision-language backbone). Trained on 127 academic decks / 3,510 slides / 25,132 component boxes over a 306-class open-vocabulary slide-component taxonomy. Mean IoU 0.873 on a slide-disjoint held-out split.

Use with the base facebook/sam3 checkpoint via the SlideForge pipeline: https://github.com/UIUC-MONET/SLIDEFORGE

Paper: SlideForge: An LLM Agent for Controllable Editing of Slides as Structured Artifacts.