InfoVis: Unified Slide Image Generation and Editing

Key points

• We formulate four slide-centric tasks and build a scalable data engine with over 13M continual-pretraining records and 1.9M supervised fine-tuning records.1
• We also introduce VisSlideBench, a benchmark of 1,527 manually screened instances that covers all four tasks with task-specific, multidimensional evaluation.1
• Using these resources, we train InfoVis, a unified 9B slide generation and editing model.1
• Starting from FLUX.2-klein-base-9B, slide-specific post-training adds 1.6 points overall, with the largest generation gains in content faithfulness; the remaining gap to proprietary models is smallest in content faithfulness and largest in aesthetic quality.1

Sources (2)

• [1]InfoVis: Unified Slide Image Generation and Editing
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 12, 01:39 PM

  Generating and editing presentation slides requires models to combine accurate text rendering, instruction following, visual grounding, aesthetic design, and precise editing within one canvas.

  We formulate four slide-centric tasks and build a scalable data engine with over 13M continual-pretraining records and 1.9M supervised fine-tuning records.

• [2]InfoVis: Unified Slide Image Generation and Editing
  Hugging Face Daily Papers · Oct 12, 12:00 AM · same content

Extractive summary: sentences quoted from the sources.