vector-genai

Sandbox for prototyping AI-driven SVG / vector graphics pipelines.

See RESOURCES.md for the survey of approaches.

Layout

samples/         input images / prompts
outputs/         generated SVGs (one subdir per pipeline)
pipelines/
  starvector/    autoregressive VLM (image → SVG), 1B/8B
  omnisvg/       autoregressive VLM with Qwen2.5-VL backbone, 4B/8B
  raster_vtrace/ flux-schnell raster → vtracer (text → image → SVG)

Each pipeline has its own uv-managed venv because the upstream repos have conflicting Python / torch requirements. Run a pipeline by cding into it and following the pipeline's own README.

Hardware target

RTX 4070 Super, 12GB VRAM. See per-pipeline READMEs for what fits.