Qualzz20

/

afv_fiber_9um

Like 0

Hi @Qualzz20 , thanks for releasing the model and the training write-up. The card notes that no independent test accuracy is claimed, so here is one measurement on a scroll this model never saw, in case it is useful for the card.

Setup. PHerc0139 (not in your training or validation scans). Five segments (w035, w039, w040, w041, w044) have a tifxyz mesh on both the 9.362 µm scan (20250728140407) and the 2.399 µm scan (20260102150214), so the same papyrus surface can be read in both. 100 tiles of about 0.9 mm. Reference: scrollprize/fiber_ink_4class_selfdistill on the 2.399 µm scan. Compared on the 9.362 µm scan: this model (revision 365e7800ac; the checkpoint is unchanged in the current e189af7507) and its base scrollprize/fiber_hz_vt (0905e68f14), both with test-time mirroring on. Measure: fiber-presence F1 per tile with a 1-pixel tolerance over a ±47 µm depth window. The null scores each map against the reference of a different tile, which keeps fiber densities but removes the spatial match.

• Difference (this model − base): −0.005 [−0.021, +0.009], so no measurable difference on this scroll in this measure. I had predicted at least +0.03 before running it.
• Per class: vertical 0.49 for both, horizontal 0.59 vs 0.58.
• Both 9 µm models mark less of the surface as fiber than the 2.4 µm model (0.30 and 0.31 vs 0.42). For this model, precision is 0.81 and recall 0.61, so what is lost is mostly fiber that is not marked.

What it is not. It is agreement with a model-made 2.4 µm reference, not accuracy. It is one scroll at 9.362 µm, so the 8.64 µm scans in your training set are untested. And it compares voxel presence maps only, so it says nothing about separating touching fibers, tracing, or the fiber volumes you build on top of the model. The fine-tune's benefit may well show there.

The protocol was committed before any model was run, and the write-up, per-tile scores and every scored map are public: write-up, raw results.

If you want it on the card, something like this would be accurate:

Agreement on an unseen scroll. On PHerc0139 (9.362 µm, 100 tiles on five segments), fiber presence agrees with a 2.399 µm reading of the same surface by fiber_ink_4class_selfdistill at F1 0.685 (0.439 with the spatial match removed; 1-pixel tolerance), about the same as fiber_hz_vt (0.691). This is agreement with a model-made reference, not accuracy. Details.

I can open a PR with that, or you can leave this as a note. Either is fine.