FervorCreative AI
Live Latest 18.11.26 · morning 401 tools tracked 992 workflows indexed 626 topics Hot: Gaussian splatting, ComfyUI, ZONOS3-Mini

3D scans have been cheap to make and impossible to fix, and Splat Repair is a free tool that lets a creator paint out a mistake and have the surface rebuild itself, which turns a ruined scan into a draft instead of a reshoot.

Splat RepairGaussian splattingLaMaPhotoshopai-editinglocal-creative-aiopen-weightscreative-workflows

Splat Repair: A Free Editor That Fixes Gaussian Splat 3D Scans Like Photoshop Fixes Photos

Scanning a room with your phone is easy now. Fixing what the scan got wrong has been close to impossible. A one-person open-source project changes that, with a brush, a clone stamp and a fill that rebuilds the wall behind the thing you erased.

Every 3D scan has something in it you did not want. A tripod leg. Your own shoe at the bottom of the frame. A sticker on a sign, a parked car, a stranger who walked through on the third pass. With a photo, you would open an editor, paint it out and move on. With a modern 3D scan, until this week, you mostly could not. You could delete the bad part and live with a hole you could see from every angle, or you could go back and shoot the whole thing again.

That asymmetry is strange when you think about it. The scanning half of the job got fast and nearly free. The cleanup half barely existed.

Splat Repair is a free editor built to close that gap. You paint over what should go, and the tool rebuilds the surface underneath, including its color and the way it catches light from different angles. Its author, Nicolas Drum, built it in eight days and published it on GitHub on November 16 under the Apache 2.0 license, which means you can use it for client work and even build on the code.

My position, after reading it closely: this is the most useful kind of AI-adjacent tool, the kind that saves a take you already liked. It is also very early, Windows-only in practice, and honest about the cases it handles badly. Both things are true.

First, what you are actually editing

Most phone scanning and 3D capture apps now produce something called a Gaussian splat. Instead of a solid surface made of triangles, the scene is stored as hundreds of thousands of tiny, soft, colored blobs floating in space. Each blob has a position, a size, a tilt, an opacity and a color that can shift slightly depending on where you view it from. Seen together, they look startlingly like a photograph you can walk around in.

That last detail is what makes splats beautiful and what makes them miserable to edit. A blob is not a pixel and it is not a polygon. It has no neighbor it is formally attached to. If you erase a sticker from a wall, there is no "wall" underneath to reveal, only a gap where blobs used to be. Fill that gap badly and it stands out from every angle except the one you fixed it from.

Splat Repair works directly on those blobs. Its working format is the common .ply file most splat tools export, and it also reads and writes .splat and .spz. The README says PLY files round-trip bit for bit, so opening and saving a scan does not slowly degrade it.

The tools will feel familiar

If you have used Photoshop, the toolbar reads like home:

  • Brush (B) paints a selection onto the 3D surface as you stroke.
  • Lasso (L) draws a freehand selection, either just the surface you see or all the way through the object.
  • Rectangle (R) for a box selection.
  • Clone Stamp (S): Alt+click to set a source, then paint to copy it somewhere else.
  • Move (G) to shift whatever you have selected.
  • Clean up, which finds "floaters" (the stray blobs that hang in midair around every scan) and lets you crop the scene with a box.

Once something is selected, you can delete it, fill it or export it on its own. The fill is where the project earns its name.

Two ways to fill a hole

Fast Fill is plain geometry, no AI model involved. It looks at the surface around the hole, works out its shape, and copies whole blobs from nearby to cover the gap. Because it copies real blobs instead of inventing new ones, their size, tilt and light behavior match the area around them. The README puts it at 0.3 to 1.5 seconds on a scan of 550,000 blobs, running on an ordinary processor. No graphics card needed.

That is the fill you will use most. It is fast enough to try, undo and try again, which is how retouching actually works.

Deep Fill is the AI route. It renders a flat picture of the scan from your current view, hands that picture to an open image-fill model called LaMa, then lifts the filled picture back into 3D blobs. This is for holes where copying nearby surface would look wrong, because the missing area needs something invented. It needs an NVIDIA graphics card, the CUDA toolkit and Microsoft's C++ build tools, about 3 GB of extra downloads, and roughly 20 seconds per fill on an RTX 4090, according to the README.

There is also a third option worth noticing. Fill from Photoshop sends the rendered view to your own copy of Photoshop and uses its Content-Aware Fill or Generative Fill. If you already pay for Photoshop and trust its fills, you can use that judgment on a 3D scan. The project is independent and not affiliated with Adobe; it drives the copy of Photoshop installed on your machine.

Where it shines

The README is refreshingly specific about this. Flat and gently curved surfaces fill in about a second: walls, signs, floors, stickers, small handles. The project's own before-and-after shows a red cross painted out of a sign on a beach hut scan.

Think about what that covers in practice:

  • Set and location scans for film previs or virtual production, where a crew member, light stand or tape mark ended up in the capture.
  • Product and prop scans where the turntable, a clamp or a fingerprint smudge is baked into the object.
  • Real estate and architecture captures with signage, clutter or a reflection you do not want on the record.
  • Game and VR environments built from scans, where a stray object would break immersion.

In each case, the value is not that the tool is magic. It is that a scan stops being all-or-nothing. A capture with one flaw becomes a draft you can finish rather than a reshoot you have to schedule.

Put this into practice

Here is a first session that will tell you within an hour whether Splat Repair belongs in your kit.

  1. Check your machine. You need Windows 10 or 11, Node.js 22 or newer, and Python 3.10 to 3.13. The README says the code is not tested on macOS or Linux, so treat those as unsupported.
  2. Install it. Clone the repository (or download the ZIP from GitHub's green Code button), open the folder and run install.cmd. Everything installs inside that one folder, and removing the tool means deleting the folder.
  3. Start it with app\start.cmd. There is an optional script that puts a shortcut on your desktop.
  4. Make a copy of a scan you already have. Work on the copy. The author says this plainly, and so do I.
  5. Run Clean up first. Floaters are the easiest win in any scan, and removing them gives you a fair picture of what is left to fix.
  6. Pick one flat problem. A sign, a sticker, a mark on a floor. Paint it with the Brush, choose Fast Fill and orbit the camera all the way around the result. A fix that only looks right from one angle is not a fix.
  7. Try one hard problem. Something with texture, like brick or carved wood. See where Fast Fill holds and where it shows. Patch what shows with the Clone Stamp.
  8. Only then consider Deep Fill. Run install.cmd deep if you have an NVIDIA card and the earlier fills could not handle a hole that needs invented detail. The README warns that the setup prints dependency conflicts along the way; it is finished when it says "Deep Fill is ready."

That order matters. Most of what goes wrong in a scan is floaters and small marks on simple surfaces, and the free, fast, no-AI tools handle those. Reach for the AI fill last.

The honest limitations

Repeating texture is hard. The README calls out strong repeating relief as the weak case, using the sole of a shoe as its example: the fill often shows and needs patching by hand. Brick, tile, wood grain, fabric weave and anything with a strong pattern will test it. Expect to clone by hand.

It is Windows in practice. Not tested on Mac or Linux. If you work on a Mac, this is a tool to watch, not one to adopt.

Deep Fill has strings attached. It needs an NVIDIA card, a compile step on first use and several gigabytes of downloads. The LaMa model's code and weights are Apache 2.0, but the README notes it was trained on the Places2 image collection, which has its own terms. If the job is commercial, read them.

A fill is a guess. Anything the tool invents, by copying or by AI, is not what was actually there. For documentation, insurance, heritage or legal work, a repaired scan is no longer a record of the place. Label it, or do not repair it.

It is days old and built by one person. The author describes it as an early release and says to expect rough edges. It had 37 stars on GitHub the morning after launch. There is no track record yet, no long list of fixed bugs, no community of users to ask.

Fixing is the new skill

For two years, the excitement around 3D capture has been about making scans faster and cheaper. That part is largely done. What separates a usable scan from a wasted afternoon now is what you can do after capture, and that has lagged badly behind.

Splat Repair is a small, rough answer to that, and it is a real one. Find the scan on your drive that you almost used but did not, because of one flaw. Make a copy, paint the flaw out and walk the camera all the way around it. If the fix holds, you just got a capture back for free. If it does not, you will know exactly what kind of damage this tool can and cannot handle, and you will shoot the next scan with that in mind.


From the briefing