Ideas worth trying that are not scheduled in ROADMAP.md. Each entry says what
we measured, what we would change and how we would know it worked. Numbers date
from 2026-08-28. Product-level ideas (and rejected ones) are in
product-ideas.md.
| Metric | Value |
|---|---|
| Bubble box F1, 147-page ground truth (student v4, laptop) | 0.971 (v3 0.968, teacher 0.967) |
| Bubble box F1, worst series (Blacksad, ground-truth boxes offset) | 0.88 (v3 and teacher 0.88); Defensores 0.98 |
| Panel box F1 | 0.93 (heuristic 0.77) |
| Uncovered area after enlarge, Ben Reilly #01 (271 bubbles) | 0.0265 page-areas |
| Uncovered area, 98-page corpus (876 bubbles) | 0.2212 |
| Uncovered area, Doom #7/#9 (78 bubbles, 6 borderless captions now boxed) | 0.0072 |
Silhouette IoU, 115 hand-validated outlines (gt/silhouettes.json) |
mean 0.881, 78/115 ≥ 0.9, 0 box fallbacks; series 0.83-0.91 |
| LaMa inpainting (spike) | 70-90 % judged good, ~4 s per crop — not shipped |
Box F1 at IoU ≥ 0.5 says whether a box exists, not whether it is tight; the Doom last-line bug was invisible to it.
115 outlines on six series, scored per series by SpeechBubbleVisualizer
(out/silhouettes.json). The worst non-scan cases are one family: the paper
body leaks into pale art beside the rim and runs to the box (Doctor Muerte
0.36/0.40, Ben Reilly 0.46, Doom 009-004 0.43-0.73). That is a new idea 3b:
stop the body at the rim ink instead of the box when the box side it reaches
is pale art, the mirror image of the Doom short-side growth.
solidPaper used a global threshold (luminance ≥ 228, or cream) tuned for
digital pages; scanner grain on 224-231 paper produced patchy bodies that
fell back to the box rectangle. PaperTone now estimates the paper per ML
box and classifies scanned paper relative to it. Spiderman 2099 fallbacks
17 → 0, IoU 0.758 → 0.879; Defensores 0.898 → 0.902; digital series
unchanged. Its two remaining weak outlines belong to idea 3b.
When the first or last line of tightly lettered text touches the bubble
outline, the paper beyond it is walled off (Ben Reilly p.17 “Y HOY N /
UMANO”, Spiderman 2099 “AAAARON”), and the copy loses part of the line.
grownTowardsShortSides handles the case where the ML box says the bubble is
taller than the body; a similar rule could close notches whose mouth lies on a
text row inside the box. Measurable against the ground truth since idea 1.
F1 0.968 → 0.971 on the ground truth grown to 147 pages; the gain is in recall (0.971 → 0.977) and on Titanes, Ben Reilly #03 and the Panini tomo. The Spiderman 2099 Vol1 04 p.7 split caption (“POR CULPA DE ESE SPIDERMAN”) renders complete on the emulator with v4; offline at 2000 px the nested second box still appears, so a merge of boxes nested in one paper body remains a cheap safety net worth adding to the outliner. v4 also boxes borderless lettering over art (Doom #7 character captions), which the outliner turns into box fallbacks — a candidate for idea 3’s family.
Uncovered area is already residual on digital pages (0.0265 / 271 bubbles). Before reaching for LaMa, try classical inpainting (border-colour average or Telea) on the thin crescents; measure on the 98-page corpus with the existing visualizer crops and only escalate to a tiny model if the crescents look wrong.
Touching bubbles, text glued to the rim and borderless sound effects are ambiguous even for a human annotator. Aim for ~95 % correct silhouettes on digital pages and ~85-90 % on scans; digital is close already.