In June 2021 the Rijksmuseum in Amsterdam showed Rembrandt's The Night Watch with its trimmed edges restored digitally: neural networks trained on Rembrandt's technique and use of color generated the missing strips, using a smaller 17th-century copy as the guide. The strips were printed and placed around the original canvas for a limited time, and the museum said it does not plan to add them to the actual restoration. The case shows what AI can contribute to art, a visual hypothesis built from evidence, and what it cannot, which is recovering what was lost.
Key takeaways
- The Night Watch was cut down in 1715 on all four sides, and the missing strips have never been found.
- The reconstruction used a copy attributed to Gerrit Lundens as its guide, plus networks trained on Rembrandt's style and colors.
- The printed extensions were shown temporarily; the museum does not intend to include them in the painting's restoration.
- A second Rijksmuseum project used AI to stitch 8,439 photographs into a 717 gigapixel image, a task of assembly, not of invention.
- Generated fill-in is an informed guess. It is not evidence of what Rembrandt painted.
What was missing and why
Rembrandt finished the painting in 1642 for the banqueting hall of the militia at the Kloveniersdoelen in Amsterdam. In 1715 it was moved to the city's town hall, now the Royal Palace, where it did not fit the intended wall. Strips were cut from all four sides, the largest from the left, and the cut-off pieces have never been found. According to the Rijksmuseum, the missing parts measure 64.4 centimeters on the left, 7 centimeters on the right, 23.3 centimeters at the top, and 11.3 centimeters at the bottom. The remaining canvas is 379.5 by 436 centimeters; the reconstruction as a whole is 393.1 by 507.4 centimeters.
What survives is a small oil-on-panel copy attributed to Gerrit Lundens and dated between about 1642 and about 1655. It measures 66.5 by 85.5 centimeters, roughly one-sixth the width of the reconstructed composition, and it is on loan to the Rijksmuseum from the National Gallery in London. Because it was made while the painting was still intact, it records the original layout, though in another hand and at another scale.
How the AI was used
The Rijksmuseum's June 23, 2021 press release says artificial neural networks were trained on Rembrandt's technique and use of color, and the system then generated the missing parts in his style. The team worked within the Operation Night Watch research project, which began in summer 2019. The release does not describe the specific models or training details, so claims about the exact architecture, training images, or error rates are not documented in the museum's own account.
The division of labor is the useful point. The Lundens copy supplied the content, meaning what stood where. The networks supplied the surface, meaning how brushwork and color might have looked at full scale. Neither source is Rembrandt's own hand for the lost areas.
What the extension changes about the picture
According to the museum, the completed composition differs from the trimmed painting in several visible ways:
- Three figures stand on a bridge at the left: two militiamen and a child.
- Captain Frans Banninck Cocq and Lieutenant Willem van Ruytenburch sit right of center instead of in the middle.
- The powder boy grasps a balustrade and has room to run ahead of the company.
- The militiaman on the far right has a complete helmet, and there is space above the ensign's standard.
The shift of the central figures is the clearest example of how a crop alters meaning. Viewers have known the trimmed version for three centuries, so the extended version reads as a different balance rather than a larger copy of the same one.
A second use: assembly, not invention
A separate Rijksmuseum project, published on January 3, 2022, produced a 717 gigapixel photograph of the painting. It combines 8,439 individual photographs, each 5.5 by 4.1 centimeters, taken with a 100 megapixel camera. Each pixel covers 5 micrometers, and the file totals 5.6 terabytes. The museum says AI stitched the smaller photographs together and that neural networks checked each photo's color and sharpness after it was taken.
The two uses differ in kind. Stitching takes images that exist and arranges them. Reconstruction produces pixels for which no image exists. The first can be checked against the photographs; the second can be judged only for plausibility.
Reading claims about AI-restored art
The table below is a general checklist built from the Night Watch example. It is illustrative, not an official standard.
| Question | Night Watch answer | Why it matters |
|---|---|---|
| What evidence guided the output? | A contemporaneous copy at a different scale | Output can be no more accurate than its reference |
| Is the result physical or digital? | Printed panels placed around the canvas | Nothing was painted onto the original |
| Is it permanent? | Shown for a limited time; not part of the planned restoration | Separates a study from a conservation treatment |
| Are methods documented? | Described in general terms only | Without details, claims of accuracy cannot be tested |
Limits and misconceptions
A network trained on one painter's style tends to produce what is typical of that style. A particular detail that Rembrandt chose, such as an unusual gesture or a color accent, is the kind of information such a system cannot supply from style alone. This follows from how generative models work in general and is a reason the museum frames the result as a view of the composition, not a recovered original.
Two misconceptions are common. One holds that AI restored the painting: the physical work was not altered by this project, and its conservation is a separate process. The other holds that AI found the lost pieces: they remain missing, and the extension is newly made. This article does not assess how the images look to viewers or whether such reconstructions are a good use of museum resources, because the sources reviewed do not settle either question.
FAQ
Was The Night Watch physically changed by AI?
No. The reconstructed sections were printed on panels and placed around the original for a limited period. The museum states it does not intend to add the lost pieces to the actual restoration.
Which painting guided the reconstruction?
A copy attributed to Gerrit Lundens, painted between about 1642 and about 1655, which shows the composition before it was trimmed in 1715.
Did the AI paint like Rembrandt?
The museum says the networks were trained on his technique and color use and generated the missing parts in his style. How closely the output matches his hand cannot be verified, because the original strips no longer exist.
Is the 717 gigapixel image generated by AI?
It is a composite of real photographs. According to the museum, AI stitched them together and checked each photo's color and sharpness.
Related reading
- Examples of AI in creative work
- Where AI and art meet
- Whether AI can be creative
- A look at AI paintings
More in AI Basics.
Last verified: October 2026
Sources
- For the first time in 300 years The Night Watch is complete again, Rijksmuseum press release, June 23, 2021
- Copy of The Night Watch (SK-C-1453), Rijksmuseum collection
- Rijksmuseum publishes 717 gigapixel photograph of The Night Watch, Rijksmuseum, January 3, 2022
- Operation Night Watch, Rijksmuseum