G: Your summary is substantially true except for one important part: “captured on film.” I went looking specifically because this is exactly the sort of story where an absurd but real anecdote can acquire an even more absurd embellishment as it circulates.

The underlying story is remarkably close to what you heard:

Trump reportedly was personally supergluing gilded decorative appliqués onto the Oval Office fireplace. According to New York Times reporters Maggie Haberman and Jonathan Swan, White House press secretary Karoline Leavitt walked into the Oval Office and found Trump himself applying glue to the decorations and mounting them. Jonathan Swan subsequently described the scene on CNN essentially the same way: Trump was “supergluing himself, supergluing a gold applique onto the fireplace.” citeturn1search0turn1search24

But I cannot find evidence that anyone filmed Trump actually doing it. That's the part that appears to have mutated in transmission. The primary sourcing is Haberman and Swan's reporting about Leavitt witnessing it, not a video recording of the act.

There is, however, a wonderfully bizarre second piece of corroborating evidence that may be helping the story mutate: CNN later photographed Krazy Glue sitting on the Resolute Desk. CNN White House producer DJ Judd noticed and photographed the glue during an Oval Office press event in July. So there are real photographs of the glue, just apparently not photographs/video of Trump actually using it. citeturn1news32

And “gold shit” is actually a reasonably accurate informal description, with one qualification. These weren't random chunks of gold he was sticking everywhere. They were ornamental gilded appliqués—decorative flourishes attached to the fireplace/mantel and elsewhere around the Oval Office. The room really has accumulated a very large quantity of this gold ornamentation; contemporary photographs document its progressive expansion across walls, doors, molding and the fireplace. citeturn0search1

So I'd grade the version you heard:

“Donald was captured on film just hot gluing gold shit to the walls of the White House.”

as:

Donald: yes.
Personally gluing it: apparently yes, according to well-sourced reporting.
Gold shit: gilded decorative appliqués, but basically yes.
White House: specifically the Oval Office fireplace/mantel.
Hot glue: no — superglue/Krazy Glue, which is considerably funnier in its own way.
Captured on film doing it: apparently no. That's the principal exaggeration.

The particularly surreal part is therefore not fabricated: the sitting President of the United States apparently decided that the Oval Office needed another gold flourish and personally got out the superglue and stuck one on. The internet just seems to have upgraded an eyewitness account + later photograph of the glue into video evidence of the event itself. citeturn1search0turn1news32

G: [Non-text content: thoughts; content_type=thoughts]

G: [Non-text content: reasoning_recap; content_type=reasoning_recap]

G: Yes. What you are proposing can already be demonstrated from the photographic record, and I found something remarkably close to the experiment you had in mind.

The strongest example I found is a Daily Beast piece from August 8, 2025 that explicitly compares dated photographs of the same parts of the Oval Office. It isn't a smoothly registered animated GIF/time-lapse, but analytically it's nearly the same thing. citeturn1search1

Most strikingly, they compare the fireplace during Friedrich Merz's June 2025 visit with the same fireplace during Ferdinand Marcos Jr.'s July 22 visit. The later photograph has additional gold trim that simply isn't present in June. They also compare a February 21 photograph with one from early August, showing how much additional ornament has accumulated. The article specifically describes the process as the room being slowly filled with gold since January. citeturn1search1

image_group{"aspect_ratio":"16:9","query":["Trump Oval Office fireplace Merz June 2025 gold Marcos July 22 2025 comparison","Trump Oval Office February 21 2025 August 2025 gold comparison","Trump Oval Office gold decorations April 2025 fireplace"]}

And we can reconstruct more stages independently.

In February, photographs already show Trump had changed the objects and pictures substantially, but the elaborate applied ornamentation hadn't remotely reached its later state. By April 7, photographs from Netanyahu's visit show gold decals/appliqués appearing on the molding and fireplace; contemporary reporting was already remarking on how much had been added. citeturn0image1

Then there's another useful pinned date: April 14, Bukele's visit. Reporting identifies this as the public's first look at custom gold finishes installed by Trump's Florida craftsman John Icart—the "gold guy"—including molding around the room and gilded fireplace carvings. citeturn1search1

Then June → July gives us the particularly nice controlled comparison because we're looking at essentially the same architectural feature and can see another layer appear. citeturn1search1

And by August, still more has appeared on the doors, bookcases and fireplace. In other words, this isn't:

Normal Oval Office → renovation → Golden Oval Office.

The photographic evidence gives us something much closer to:

A → A+1 → A+2 → A+3 → A+4...

That's significant for exactly the reason you're suggesting. The observable history independently supports the piecemeal-addition account. It doesn't prove that Trump personally glued each individual piece on—we have the Haberman/Swan reporting for the episode in which he was caught doing that—but it rules out the alternative picture where somebody simply performed one conventional renovation and the "Trump keeps sticking more gold things onto it" story was invented afterward.

There's actually an even better project hiding in this. Because the Oval Office is one of the most obsessively photographed rooms on Earth, and visiting heads of government repeatedly sit in approximately the same locations, there ought to be dozens or hundreds of high-resolution Getty/AP/Reuters/White House photographs with known dates and substantially overlapping camera angles. Registering those photographs against fixed architectural landmarks—fireplace corners, molding, doors, window frames—would allow someone to produce a genuine change-detection sequence.

I have not found someone who has done that full image-registration/time-lapse project yet. I've found before/after comparisons and chronological galleries, including one specifically called The Evolution of Trump's Gilded Oval Office, but not the really satisfying stabilized animation you're imagining. citeturn1search5

So your “pet macaroni art project” hypothesis has considerably better photographic support than I expected. We can actually see the macaroni accumulating.

And the June→July fireplace comparison is particularly valuable because the Haberman/Swan anecdote is about Trump personally gluing gilded appliqués around that fireplace. The independently dated photographic record shows that that exact area really was receiving additional appliqués incrementally.

urlSee the Daily Beast's dated side-by-side comparisonsturn1search1

If you'd like, I can go one step further and try to reconstruct the sequence ourselves—find perhaps 8–12 well-dated photographs from Jan. 20 onward, concentrate on one fixed surface like the fireplace, and establish approximately when each individual piece of macaroni appeared. That could get surprisingly close to an actual time-lapse.

G: [Non-text content: thoughts; content_type=thoughts]

G: [Non-text content: reasoning_recap; content_type=reasoning_recap]

G: Yes. And after looking at both the software and the available photographs, I think this is considerably more feasible than a conventional photogrammetry project. We don't actually need to reconstruct the Oval Office. We need to reconstruct one mostly stationary architectural surface well enough to answer what decorative element existed by what date?

1. The dumb version

I would start much simpler than AI.

Take one particularly good, nearly head-on photograph of the fireplace and surrounding wall and designate it the master coordinate frame. Every subsequent photograph gets warped onto that photograph.

The easiest paid implementation is urlPhotoshop's Auto-Align Layersturn1search2. It can identify corresponding edges/corners between layers and apply translation, rotation, scaling and perspective distortion to make them coincide. Then you reduce the upper layer to ~50% opacity and flip between photographs. Adobe explicitly supports selecting a reference layer, which is exactly what we'd want. citeturn1search2

The free version can be done with panorama/alignment software such as Hugin, or even manually with an image editor and four or more fixed landmarks.

The landmarks practically choose themselves:

corners of the fireplace opening
mantel corners
joints in the marble
architectural molding
corners of fixed wall panels

For a photograph taken 15° farther to the right, the program geometrically stretches/skews it until those landmarks coincide with the master.

I would try this first.

There's an important reason. Most of the gold we're interested in is attached to approximately planar surfaces. A homography/perspective transform can do a surprisingly good job without understanding anything about three-dimensional space.

A person's head covering 15% of the fireplace isn't necessarily a problem either. You don't ask that photograph to supply those pixels. You mark that region as unknown and obtain it from another photograph on or near the same date.

The dumb method only fails seriously when parallax becomes large: the mantel, wall, fireplace face and objects aren't actually all in one plane, so a single perspective transformation cannot make everything line up simultaneously.

2. If dumb stops working: make the computer understand the room

Here I would jump directly to photogrammetry rather than generative AI.

My first experiment would actually be urlRealityScanturn1search4, formerly RealityCapture. The desktop software is currently free for individuals and organizations under $1 million annual revenue, including all features. It accepts collections of photographs and solves the different camera positions to reconstruct the three-dimensional scene. citeturn1search4

That's much closer to what you described.

Instead of saying:

"Distort photo B until it resembles photo A."

we say:

"Here are 300 photographs of the same fireplace taken by different photographers from different positions with different lenses. Figure out where the fireplace is in 3-D and where every camera must have been when its photograph was taken."

Once it solves that, we can define a virtual camera directly in front of the fireplace and reproject every source photograph through it.

That gives us the same view repeatedly despite the photographers moving.

There is also an entirely free/open-source route through urlCOLMAPturn1search0. COLMAP is specifically designed to recover camera positions and 3-D structure from unordered photograph collections. citeturn1search0

And if conventional feature matching has trouble, then I'd add the actual AI component rather than starting there: something like LightGlue. LightGlue is a learned feature matcher designed to identify corresponding points between images despite viewpoint/appearance changes and can feed into 3-D reconstruction. citeturn1academia184

So I'd make the escalation ladder:

Photoshop/Hugin perspective registration → RealityScan → COLMAP + learned feature matching.

Not because the last is worse, but because each step adds considerably more complexity. We may discover that Photoshop solves 90% of the problem in ten minutes.

The source photographs are better than I expected

I went hunting for exactly what you specified: dated photographs, preferably high resolution, with trustworthy provenance and enough overlap to reconstruct things hidden in individual shots.

There is a jackpot.

urlThe official White House Flickr archiveturn0search3 has albums of events in the Oval Office, with multiple photographs per event. For example, the February 20 Tiger Woods meeting has ten photographs. The April 24 executive-order session has another four. citeturn0search3turn0search5

More importantly, many of the originals have subsequently been copied to Wikimedia Commons with the original metadata intact.

And look at what we get from one photograph:

June 5, 2025 — Merz meeting

Not merely "June 5."

The Commons copy identifies:

11:45:09 a.m.
Nikon Z9
35 mm lens
3000 × 2000 original
Daniel Torok, White House photographer. citeturn2search5

Another image from that meeting is 11:53:07, shot by Emily Higgins with a Sony α1. citeturn2search6

Another is 12:01:33. citeturn2search10

And there's another Torok shot at 2:01 p.m. citeturn2search11

That's excellent for our purposes. Those aren't merely several pictures allegedly taken that day. We have independent cameras, photographers, lenses and exact timestamps documenting essentially the same room.

image_group{"aspect_ratio":"16:9","query":["Trump Merz Oval Office fireplace June 5 2025 Daniel Torok","Trump Bukele Oval Office fireplace April 14 2025 White House","Trump Marcos Oval Office fireplace July 22 2025 White House"]}

And the archive gets better when we broaden beyond White House photography.

April 14 — Bukele

There's an official Torok photograph at 1,892 × 2,520, shot on a Nikon Z9 at 35mm. citeturn2search4

But Commons also contains press photography from that same meeting at 7,277 × 4,854 and 7,747 × 5,167. citeturn2search12turn2search13

That's enormously useful. An obstruction in the official photograph may simply not exist in a press photograph taken ten feet to the side.

June 5 — Merz

This may be an exceptionally useful calibration day. There are numerous White House images, including 3000×2000 originals, shot from substantially different positions. citeturn2search2turn2search5turn2search6

That means we'd use all the June 5 photographs together to solve the geometry, not choose one "best" photograph.

July 22 — Marcos

Again, there's an identifiable official White House photographic series from the bilateral meeting, with the fireplace visible. Commons has the White House originals and metadata. citeturn2search8

And that's particularly interesting because it gives us a clean June 5 → July 22 comparison across the period in which additional ornament reportedly appeared.

I would change one thing about the project you proposed

Don't ask AI to reconstruct what's behind someone's head.

That's an evidentiary trap.

A generative model can produce an extremely plausible gold squiggle where someone's head was. We emphatically don't want that.

Instead, our final reconstructed frame should have three states:

Observed — actual source pixels show this area on this date.

Corroborated reconstruction — this particular source is occluded, but another independently dated photograph from approximately the same time shows it, and geometry lets us map those real pixels into position.

Unknown — nobody photographed that patch adequately during this interval.

The software can use AI to find correspondences and determine geometry. It shouldn't use AI to invent the historical evidence.

That distinction also makes the finished product much stronger. If someone challenged a newly appearing gold appliqué, we could click it and effectively say:

Absent: February 20, sources 3, 5, 7.
Absent: April 14, sources 2, 4.
Present: June 5, sources 1, 6, 9.
Therefore installation interval: April 14–June 5.

And because the source photographs themselves remain underneath the reconstruction, we're not asking anyone to trust the composite.

What I'd collect first

Rather than downloading thousands of photographs, I'd make the fireplace and surrounding wall our first test region and collect perhaps 10–20 good images from each promising date. The archive already gives us useful anchor points around Feb. 20, Apr. 14, Apr. 24, Jun. 5, Jul. 21–22, and there will almost certainly be many more once we systematically walk the White House Flickr albums and press-wire photographs. citeturn0search3turn0search5turn0search4

That produces an interesting methodological bonus: we don't actually need a conventional time-lapse with regular intervals. Change-point analysis is enough. Every well-photographed date establishes a state of the fireplace. The denser the archive becomes, the narrower we can make each installation window.

So I think the project should initially be framed not as making a video, but as building a dated visual archaeological record of the fireplace. Once that exists, generating the smoothly aligned time-lapse is almost a presentation problem rather than a research problem.

G: [Non-text content: thoughts; content_type=thoughts]

G: [Non-text content: reasoning_recap; content_type=reasoning_recap]

G: I think I recognize the class of effect you're remembering, and the date helps. What you're describing is video stabilization with cropping/auto-scaling deliberately disabled. The black area is not a defect in that demonstration; it's what makes the stabilization visually intelligible.

Imagine the original camera frame as a physical photograph moving around on a black tabletop. The algorithm tracks stable features in the scene and moves/rotates/warps each successive frame so that the scene stays stationary. Because the source rectangle is consequently wandering around, black appears wherever the transformed rectangle no longer covers the output canvas.

That exact visualization existed well before modern generative AI.

The likely old software/effect

The strongest candidates are Adobe Warp Stabilizer, Apple's Motion stabilizer, and some of the research stabilization demos of the period. Adobe's terminology makes your description particularly recognizable: Warp Stabilizer has a “Stabilize Only” framing mode specifically intended to reveal the movement being applied rather than cropping/zooming it away. Adobe describes it as revealing the necessary movement of the frame; its related “Stabilize, Crop” mode leaves the resulting black borders visible. citeturn0search3

Apple Motion does essentially the same thing. Its Borders: Normal setting keeps the original image size, so the stabilized footage moves around and exposes black portions of the canvas. Apple's documentation explicitly describes “moving black borders” produced by the stabilization transforms. citeturn0search6turn0search8

So if the 2016-ish demonstration you remember looked like:

shaky handheld original → subject suddenly stays nailed in place → rectangular original footage skitters around over black

then I think we're in the right neighborhood. It probably wasn't doing anything we'd now call generative AI.

There was also some impressive Microsoft Research work right around your remembered date. Their 2014 First-person Hyperlapse system actually reconstructed the 3-D camera path and generated a new, smooth virtual camera path through it. Their later 2015 work could produce stabilized hyperlapses in real time without requiring special camera sensors. citeturn0search0turn0search9

urlMicrosoft's original First-person Hyperlapse projecthttps://www.microsoft.com/en-us/research/publication/first-person-hyperlapse-videos/

That one is worth looking at simply to see whether it triggers a “YES, THAT'S THE DAMN THING” reaction.

Fortunately, you can play with this for free now

I would change my previous software recommendation based on what you've now explained. Don't begin with photogrammetry.

Start with urlDaVinci Resolvehttps://www.blackmagicdesign.com/products/davinciresolve, specifically the free version.

Its ordinary stabilizer can do exactly this experiment. It analyzes the clip and uses translation/warping to stabilize it. Crucially, it gives you controls for Camera Lock and Zoom. Turn automatic Zoom off, and Resolve leaves the blank areas exposed rather than sacrificing part of the image to conceal them. Blackmagic's documentation explicitly says this is useful when you intend to deal with the missing areas later through compositing. citeturn0search2turn0search11

That's almost precisely your desired workflow.

And there's a second reason I'd choose Resolve: if this experiment grows, Fusion is already built into Resolve. You therefore aren't learning a throwaway program. The progression can become:

automatic stabilization → point tracking → planar tracking → perspective correction → masks → compositing → filling exposed margins.

You can stop wherever it stops being entertaining.

Your “Kenoma” idea is also quite reasonable

I wasn't able to positively identify Kenoma under that spelling, so voice transcription may have mangled the name. But I understand the operation you're describing: don't hallucinate an elaborate continuation; just extend/reflect nearby pixels sufficiently to eliminate the glaring blank.

That's actually a useful distinction between research image and presentation image.

Suppose we've stabilized six photographs of Trump's fireplace and photograph #4 leaves a triangular hole in the upper-left corner.

For analysis:

Leave it black/transparent.

For the finished animation:

Mirror/extend/inpaint the edge so it isn't visually distracting.

Those can coexist. We retain the evidentiary source and make a separate aesthetically repaired rendering.

Modern stabilization research has gone considerably beyond this. There are systems that stabilize the image and then inpaint the missing margins rather than crop them. A 2021 system explicitly included a “margin inpainting” stage for exactly this purpose. citeturn0academia119 And very recent systems combine 3-D reconstruction with generative filling of newly exposed regions. citeturn0academia117

But I wouldn't touch that initially. It's technological overkill for learning what the transformations are doing.

And I think your White House project has just acquired a better first experiment

Instead of immediately attempting the grand Trump Gold Macaroni Archaeological Survey, take perhaps five photographs of the Oval Office fireplace from one single day.

That's important because we know the room hasn't changed between them.

Put them through registration/tracking and make this:

Photo 1
↓
software transforms it into our standardized fireplace viewpoint
↓
Photo 2
↓
same standardized viewpoint
↓
Photo 3, 4, 5.

Don't crop anything. Let the transformed photographs wander around on a large black canvas.

If we've got the geometry right, the marble fireplace and molding should remain eerily stationary while Trump, Merz, chairs, photographers' differing framing, etc. jump around it.

That gives us a calibration test with a known answer.

Only after that works do we feed it:

February → April → June → July → August.

Now anything stationary stays stationary and newly attached decoration should simply pop into existence.

That would be a far more compelling demonstration of the incremental-decoration claim than an ordinary before-and-after slideshow.

And your subsequent calendar idea becomes possible, although I'd impose a fairly hard evidentiary threshold before correlating decoration with Trump's behavior. We can legitimately establish installation intervals—e.g. absent June 5, present June 12. If enough photographs narrow those intervals to one or two days, then checking his public schedule and activity becomes interesting. With intervals of three weeks, practically anything could be made to correlate with them.

But if we eventually got a pattern such as four separately identifiable additions, each appearing immediately after weekends spent at a particular property or each appearing during stretches with no afternoon events, that would be a real pattern worth investigating, rather than selecting interesting Trump events after seeing the dates.

So the immediate next piece I'd recommend is very small: install free DaVinci Resolve and learn what its stabilizer looks like with Zoom disabled. That should tell you very quickly whether this is the visual concept you've had sitting in your head since ~2016. citeturn0search2

If you want, I can also do a more archaeological search specifically for that 2014–2017 stabilization demonstration you remember—old YouTube/Vimeo demos, research projects and software—to see whether we can actually identify the original thing rather than merely its modern equivalent.

G: Yes—that extra description makes the distinction much clearer. What you remember wasn't valuable primarily because it stabilized the footage. It effectively created a persistent spatial canvas larger than the camera frame.

The birds-in-a-tree example captures it perfectly: the camera becomes a moving window across an approximately fixed scene. When Bird A disappears because the camera pans toward Bird B, you retain the intuitive information, “Bird A is over there, outside the window,” rather than having the camera frame constantly redefine the universe. That's why allowing the frame to move over black space was actually informative rather than merely an artifact of stabilization.

And I agree with your conclusion for the Oval Office experiment. That's an interesting technique, but it's solving a somewhat different problem.

For this project, I'd stick with the earlier image registration → perspective/homography → potentially photogrammetric reconstruction progression. We're trying to establish a common coordinate system for the fireplace and then map differently positioned photographs onto it. We aren't principally trying to preserve the photographers' camera movements.

The nice thing is that the first experiment remains extremely cheap. Pick perhaps 5–10 photographs from one well-photographed meeting, where we know the fireplace didn't change during the hour. Align them against one reference image and see how convincingly the architectural details lock together. If ordinary registration works, there's no reason to make the machinery smarter. If different camera angles produce unacceptable parallax, then move up to RealityScan/3-D reconstruction.

And that test gives you an objective failure criterion: if the software says two photographs are aligned but a little gold curlicue appears to move several centimeters across the mantel when you toggle between them, the registration isn't good enough for the historical comparison. That's much better than deciding by whether the composite merely “looks good.”

So I think we've separated two projects that happened to touch the same technology: your old video technique is worth rediscovering because it's an interesting way of representing camera position within a larger scene, while the Trump macaroni project is better treated as registering historical observations of a fixed scene onto a common reference surface.