If you’ve tried plotting AI-generated images, you probably know the struggle: generic tracing tools (like Potrace) trace the outline of a line, resulting in double-strokes that ruin the look and take twice as long to plot.
What I tried previously:
- Potrace / Inkscape Trace: Great for filled shapes, but results in "hollow" lines for line art.
- Canny Edge Detection: Often too messy; it picks up noise and creates jittery paths.
- Standard SDXL: Struggled with geometric coherence, often breaking lines or hallucinating perspective.
- A bunch of projects that claimed to be txt2svg but which produced extremely poor results, at least for pen plotting. (Chat2SVG, StarVector, OmniSVG, DeepSVG, SVG-VAE, VectorFusion, DiffSketcher, SVGDreamer, SVGDreamer++, NeuralSVG, SVGFusion, VectorWeaver, SwiftSketch, CLIPasso, CLIPDraw, InternSVG)
My Approach:
I ended up writing a Python tool that combines a few specific technologies to get a true "centerline" vector:
1. Prompt Engineering: An LLM rewrites the prompt to enforce a "Technical Drawing" style optimized for the generator.
2. Generation: I'm using Flux.2-dev (4-bit). It seems significantly better than SDXL at maintaining straight lines and coherent geometry.
3. Skeletonization: This is the key part. Instead of tracing contours, I use Lee’s Method (via scikit-image) to erode the image down to a 1-pixel wide skeleton. This recovers the actual stroke path.
4. Graph Conversion: The pixel skeleton is converted into a graph to identify nodes and edges, pruning out small artifacts/noise.
5. Optimization: Finally, I feed it into vpype to merge segments and sort the paths (TSP) so the plotter isn't jumping around constantly.
You can see the results in the examples inside the Github repo.
The project is currently quite barebones, but it produces better results than other options I've tested so I'm publishing it. I'm interested in implementing better pre/post processing, API-based generation, and identifying shapes for cross-hatching.