Supplementary figures for our response. Each figure lists the reviewer questions it answers; quantitative results are in the tables of the response.


Figure R1. Oracle vs. automatically extracted observed SVG

Reviewer tPTt Q1 · Reviewer iHcn W4 / Q4 · Reviewer 2Fi9 W1 / Q1

image.png

Both settings use the same damage mask. The extracted paths match the oracle ones; along the mask boundary they are smoother, because Potrace fits the cut instead of following pixel steps.

On 1,600 test instances, Mask IoU is 81.23% with oracle and 79.56% with extracted observed SVG (Table R1).


Figure R2. Restoration of naturally damaged manuscripts

Restoration examples using observed SVGs extracted directly from damaged scans and automatically estimated, manually refined damage masks.

image.png


Figure R3. Direct binarization and vectorization

Otsu binarization followed by Potrace on the same manuscript examples, showing residual background noise and fragmented strokes.

image.png

Figure R4. General-purpose MLLM completion

Qwen3.8-Max results in fast and thinking modes, composed with the same observed SVGs using only predictions inside the damage mask.

image.png


Prompt for the MLLM baseline (Table R3)