GCH 2026 · ACM JOCCH Special Issue on Visual Heritage

TEXMET: A Curated Dataset and Deep Learning Benchmark for Archaeological Textile Reconstruction

Hamza Zafar · Jon Yngve Hardeberg · Davit Gigilashvili

Colourlab, Department of Computer Science, NTNU

Oseberg fragment: G. A. Pantos, Museum of Cultural History, UiO

Oseberg ship burial · Norway · 834 CE

A Viking ship grave of two high-status women, found in 1903. Its textiles are among the finest of the Viking Age, and survive only as fragments.

Hundreds of pieces.

No ground truth.Nobody knows how they originally fit together.

Goal: put them back together.

Oseberg ship burial, 834 CE. Images: G. A. Pantos, Museum of Cultural History, UiO. Not to scale.

What we have to work with

Fadedcolors almost gone

Torn edgesirregular, not a grid

Missing partsholes, lost areas

Today's puzzle benchmarks look nothing like this.

Oseberg fragment, 834 CE. Image: G. A. Pantos, Museum of Cultural History, UiO.

Can AI put the pieces back?

Illustration · 25 pieces, shuffled and rotated

Puzzle solver

An AI model: give it shuffled pieces, it predicts where each one goes.

Today's best solvers are strong on clean photos.

But do they work on a textile nearly 1,200 years old?

Allori, The Gathering of Manna, 1595–96. The Met, Open Access (CC0); in the TEXMET test set.

From the real problem to a first step

The real problem

Which piece fits where?

Our first step

Can a model read the weave well enough to put it back?

Standard benchmark setting: 9 pieces, no rotation. Each Oseberg fragment photo becomes its own puzzle.

Oseberg fragments: G. A. Pantos, Museum of Cultural History, UiO. Not to scale.

Three state-of-the-art solvers

JPDVTCVPR 2024

Starts from a noisy guess of where pieces go, and cleans it up step by step.

FCViT2025

Predicts the coordinates of each piece directly.

DiffAssembleCVPR 2024

Links the pieces in a graph: each piece looks at its neighbors.

All three learned from ImageNet: 1.28 million everyday photos.

Animations are illustrations of each idea. Liu et al. CVPR 2024; Kim et al. 2025; Scarpellini et al. CVPR 2024.

Why a new dataset?

All three learned from everyday photos.
Not from historical textiles.

We looked for textile data

✗

TextileNet

760k images, but modern fashion. Nothing historical.

✗

RePAIR

Real fragments, but frescoes, not woven textiles. Restricted license.

✗

Puzzle benchmarks

ImageNet, CelebA-HQ: everyday photos and faces. No textiles.

Images: TextileNet, Zhong et al. 2023, Fig. 4; RePAIR, Tsesmelis et al., NeurIPS 2024, Fig. 1; ImageNet samples.

Nothing fit. So we built one.

Building TEXMET

−4,908Poor quality
armor, weapons

−3,808Noise
threads, near-empty, incl. 2,567 duplicates

55,007

train 14,915val 1,864test 1,865

Real TEXMET candidates and flagged images from our curation sessions (The Met Open Access, CC0). Map and fingerprints are schematic.

Introducing TEXMET

18,644

historical textiles, curated for puzzle solving

European 46%

Asian 28%

Islamic 16%

11 more

CC0

free for anyone

1,000+

downloads

14,915 / 1,864 / 1,865

train / val / test

QR: huggingface.co/datasets/hzafar/TEXMET

Get the dataset

Images: The Met Open Access (CC0). Share of 18,644 by Met department. Downloads: Hugging Face, Oct 2026.

Round 1: the solvers as they are

Trained on ImageNet only, then tested on TEXMET 1,865 images and Oseberg 45 real fragments

FCViT collapses68.8% → 8.3% on TEXMET

JPDVT halves67.9% → 32.9% on TEXMET

DiffAssemble holds up56.6% → 65.1% on TEXMET

On real Oseberg, all three ≤ 27%JPDVT 20.0 · FCViT 24.4 · DiffAssemble 26.7

Puzzle accuracy: share of puzzles with every piece in the right place. 3×3, no rotation.

They struggle on textiles

Does learning from TEXMET help?

We fine-tuned all three on TEXMET 14,915 training images.

JPDVT

Catastrophic forgetting

32.9 → 10.9%on TEXMET

FCViT

Barely adapts

8.3 → 10.9%on TEXMET

DiffAssemble

Adapts. We carry it forward. It also handles:

rotation

missing pieces

any puzzle size

→

Oseberg is never used for training.

DiffAssemble: before → after TEXMET

2.4×

on real Oseberg fragments

12 → 29 of 45 fully reassembled.
Never seen in training.

Puzzle accuracy, 3×3. TEXMET test: 1,865 images. Oseberg: 45 fragments.

DiffAssemble on real Oseberg fragments

9/9 ✓

Every piece back in place.

Each fragment photo cut into 3×3, shuffled, and reassembled by the model.

DiffAssemble reconstructions of Oseberg fragment photos. Images: G. A. Pantos, Museum of Cultural History, UiO.

Is it reading the fabric, or just the outline?

Fabric only: square crops from inside each fragment, no outline.
206 crops from the 45 Oseberg fragments.

Outlines help. But on fabric alone, the fine-tuned model still wins: 39.8 vs 27.7.

DiffAssemble on Oseberg, 3×3 puzzle accuracy. Random guessing ≈ 0.0003%. Image: G. A. Pantos, Museum of Cultural History, UiO.

Real fragments have holes.

Illustration

1 · When parts are missing, do the solvers still work?

Not all of them. FCViT falls from 70.6% to 46.1%. DiffAssemble barely moves.

2 · Can an AI that fills in the holes rescue them?

Partly. Every model improves, none gets back to where it started.

DiffAssemble, the model we carried forward, copes best with holes.

Test: 1,000 ImageNet photos, 3×3 puzzle accuracy, mask from one Oseberg fragment (35.6% removed), inpainting by ASUKA. Illustration: A. Allori, The Gathering of Manna, The Met (CC0), not a test image.

Real fragments are never complete: holes, frayed edges. Two questions.
This shape is the outline of one real Oseberg fragment, holes included. We used it as a stencil.
[click] Here on a tapestry to show the idea. The real test: 1,000 complete everyday photos (ImageNet), because to measure damage you need the complete picture. About a third of each photo is gone. Question 1: do the solvers still work? Not all of them. FCViT drops from 71 to 46. DiffAssemble barely moves.
[click] Question 2: can an AI that fills in the holes rescue them? We used ASUKA, an inpainting model. Partly: everyone improves, nobody fully recovers.
[click] And the model we carried forward, DiffAssemble, is the one that copes best.
(If asked why photos, not textiles: a controlled test, same damage on every image. Paper also tests a plain rectangle mask: shape matters more than amount.)
(If asked: JPDVT scores 87.8 on this 1,000-photo subset vs 67.9 on the full split; compare within this test only.)

Limitations and future work

Limitations

Future work

Regular 3×3 grids, no rotationevery piece is a square of the same size

→

Irregular fragment geometriesreal shapes, any size, with rotation

Within-fragment onlyno ground truth for how fragments fit together

→

Multi-textile reconstructionwhole textiles from mixed fragment collections

45 fragments, synthetic cutsTEXMET puzzles are cut from intact images

→

Larger, real fragment datasetsincl. real archaeological fragment imagery

Beyond textiles: the same approach may carry over to ceramics, frescoes and manuscripts.

Key takeaways

18,644

TEXMET: the first textile benchmark for puzzle solving

Curated from The Met Open Access, public on Hugging Face.

26.7 → 64.4%

Domain adaptation is essential

Natural-image models fail on textiles. Fine-tuning DiffAssemble on TEXMET, tested on Oseberg.

39.8% vs 27.7%

Textile features, not just boundaries

Fine-tuned vs baseline on inscribed, fabric-only crops of Oseberg fragments.

Thank you.

Questions?

Now recruiting: 15 PhD positionsUNCHANGE, an MSCA Doctoral Network in computational and heritage science, 2027–2031. Scan the code on the flyer.