TLDR: A training-free AI system called “See it. Say it. Sorted.” uses a Vision-Language Model (VLM) and Large Language Models (LLMs) in an iterative Critic-Candidates-Judge loop to convert hand-drawn sketches into precise, editable Scalable Vector Graphics (SVG) diagrams. It outperforms other advanced image generation models by focusing on qualitative reasoning and producing structured outputs, making it adaptable for various design applications.
Imagine being able to quickly turn a rough sketch on a napkin into a perfectly aligned, editable digital diagram. A new research paper introduces an innovative AI system called “See it. Say it. Sorted.” that does exactly this, transforming hand-drawn sketches into precise, compositional diagrams like flowcharts.
Traditional methods, especially those based on diffusion models, are excellent at creating photorealistic images but often struggle with the exactness and structured nature required for diagrams. They find it hard to maintain distinct shapes, lines, and perfect alignment, and aren’t easily adapted to user instructions or varying diagram sizes.
The “See it. Say it. Sorted.” system takes a different approach. It’s a training-free agentic system that combines the power of Vision-Language Models (VLMs) with Large Language Models (LLMs) to produce editable Scalable Vector Graphics (SVG) programs. This means the output isn’t just a picture; it’s a structured code that can be easily modified and integrated into design tools like PowerPoint or Google Slides.
How Does It Work? The Critic–Candidates–Judge Loop
The core of this system is an iterative process that mimics how a human might refine a drawing. It operates in a loop involving three main components:
- Critic VLM: This component acts like an art critic. It compares the original hand sketch with the current version of the digital diagram. Instead of giving exact numbers, it identifies a few key discrepancies and suggests qualitative, relational edits. For example, it might say, “The blue rectangle should just touch the red circle on its left,” rather than giving precise coordinates. This qualitative feedback is crucial because VLMs are better at understanding relationships than exact measurements, and it gives the LLM more flexibility.
- Candidate LLMs: Based on the Critic’s feedback, multiple LLMs (Large Language Models) generate several different SVG modifications. These candidates explore various strategies, from conservative adjustments to more aggressive changes, ensuring a diverse set of options.
- Judge VLM: Finally, a Judge VLM evaluates all the proposed candidates, along with the current diagram, and selects the one that best matches the original sketch. If no candidate improves the result, the system can revert and the Critic VLM gets feedback to refine its suggestions. This ensures stable and continuous improvement towards the target diagram.
This design prioritizes qualitative reasoning, which helps maintain global constraints like alignment and connectivity, and naturally supports human-in-the-loop corrections. The system is also highly adaptable, handling different sizes and complexities without needing to be retrained, and its outputs are interpretable and object-based SVG code.
Also Read:
- Unlocking Analog Circuit Design for AI: Introducing Image2Net
- LENS: Unifying Language Understanding and Pixel Segmentation
Impressive Results
The researchers evaluated “See it. Say it. Sorted.” on 10 sketches inspired by flowcharts from published papers. The system consistently outperformed two leading closed-source image generation models, GPT-5 and Gemini-2.5-Pro. While those models often missed structural details or inserted unwanted text, “See it. Say it. Sorted.” faithfully reconstructed layouts, maintained precise alignments, and accurately composed complex elements like multi-headed arrows.
The paper highlights that because the outputs are programmatic SVGs, the approach can be easily extended to various presentation tools via APIs. As VLMs and LLMs continue to improve, the accuracy and efficiency of this system are expected to grow without requiring any retraining. The codebase for “See it. Say it. Sorted.” is also open-sourced, allowing others to explore and build upon this exciting development. You can find the full research paper here: See it. Say it. Sorted: Agentic System for Compositional Diagram Generation.
