Artifact: Pelican Bike SVG
Public page: /artifacts/gpt-image-2/pelican-bike-svg.html
Original raster: /artifacts/gpt-image-2/pelican-bike-original.png

Model attribution:
- Underlying image model: gpt-image-2
- Chat/model driver used by term-llm: ChatGPT gpt-5.4-mini
- Jarvis generated the raster through term-llm's configured image provider, then vectorized it locally.

Prompt used for the original image:
A whimsical high-quality illustration of a pelican riding a bicycle along a sunny seaside path. The pelican is balanced confidently on the bike, wings lightly on the handlebars, big beak, playful expression, ocean and blue sky in the background. Charming, detailed, vibrant colors, slightly surreal but believable, no text.

Original raster:
- File: pelican-bike-original.png
- Dimensions: about 1254 x 1254 px
- Size at generation: about 2.5 MiB

Vectorization goal:
Convert the generated PNG into a shareable SVG under 3 MiB without using an embedded raster image. The final artifact should remain recognizably close to the source image, but be represented as SVG paths.

Tools installed/used locally:
- ImageMagick: preprocessing, resize/posterization, preview rendering
- vtracer 0.6.5: color raster-to-SVG path tracing
- scour: SVG cleanup/minification
- potrace, Pillow, OpenCV were installed while exploring the pipeline, but the final keeper used ImageMagick + vtracer + scour.

Final pipeline:
1. Start from the generated PNG.
2. Resize/posterize with ImageMagick to reduce high-frequency painterly noise that would explode path count.
3. Run vtracer in color/spline/poster mode to trace color regions into SVG paths.
4. Use stacked hierarchical tracing so shapes layer naturally instead of leaving too many cut-out gaps.
5. Tune speckle filtering and path precision to keep the image under 3 MiB while retaining the pelican, bike, beach, and coastal background detail.
6. Run scour to remove comments/IDs/metadata and compact the SVG.
7. Verify the SVG is pure vector by checking that it contains no <image> tag.
8. Render the SVG back to PNG with ImageMagick as a sanity preview.

Final vtracer-style settings used:
- preset: poster
- mode: spline
- hierarchical: stacked
- color_precision: 6
- filter_speckle: 14
- path_precision: 1
- corner_threshold: 60
- segment_length: 8
- splice_threshold: 45

Final SVG:
- File: pelican-bike-svg.html contains the inline SVG artifact page
- Raw inline SVG payload size before HTML wrapping: 1.81 MiB
- Page size: about 1.9 MiB
- Contains no embedded raster image tag

A rejected/alternate approach:
Before using vtracer, Jarvis wrote a custom Python quadtree vectorizer using ffmpeg as the decoder. It converted the image into thousands of colored SVG rectangles. That version was only about 330 KiB and had a nice mosaic/poster charm, but vtracer produced the better shareable artifact.