"AI retopology in seconds" is a real feature and it does what it says on the button. It is also not retopology, and if you intend to bake a high-poly onto the result, that distinction is the whole job.
I sell forty base mesh packs and I have spent twenty years cleaning up geometry so that a projection lands. So when the remesh buttons started appearing inside every generator, the first thing I did was not look at the wireframe screenshot. I looked at what happens when you bake a sculpt onto the mesh they hand back. That is the test that separates a tool from a demo, and almost nobody publishes it, because almost everybody writing about this is selling credits.
The vendors are not lying. They are answering a different question.
Read the documentation instead of the landing page and the answer is already sitting there.
Meshy's own AI Retopology Guide says the output is either a quad-dominant mesh or a lean triangle mesh, that the target ranges from 100 up to 300,000 polygons, and — the sentence everyone scrolls past — that remeshing rebuilds the polygon structure, which invalidates the old UV layout, so textures do not survive the trip. The same page says that for a hero character headed into close-up film shots, artists still do a manual pass to steer edge loops around the eyes and mouth.
3D AI Studio's Remesh page says its job is to reduce polygon counts and to convert messy triangle meshes into clean quad topology. Its texture option rebakes the original surface appearance onto the new mesh topology, at up to 4096 pixels.
Read that last one slowly. It rebakes appearance — the colour that was already on the model. It is not projecting a sculpt into a tangent-space normal map through a cage you control. Those are two different operations wearing the same verb.
So the honest summary of both products is that they reduce polygons and rebuild a surface. Retopology, the way an environment or character artist uses the word, is deciding where the polygons go. Neither tool claims to do that, and both of them say so in writing.
The test that matters is the bake, not the wireframe
If you want to know whether any of this works in your pipeline, do this once, on an asset you have already shipped.
- Take a sculpt you have a known-good hand-built low-poly for.
- Remesh it at the vendor's own recommended budget, not at one that flatters the result. 3D AI Studio says game assets typically need 5,000 to 15,000 faces. Meshy suggests around 50K faces for a PC model and about 10K for mobile.
- Auto-unwrap and auto-repack the output. Do not hand-fix anything. You are testing the tool, not your patience.
- Bake normal, curvature, ambient occlusion and world-space position from the same high-poly onto both low-polys. Same cage, same ray distance, same 4K output, same 4px padding.
- Look at all four maps on their own, side by side, before you open a single material.
Step five is the one people skip, and it is where the entire answer is. You are not comparing renders. You are comparing whether the surface data survived the projection.
The density is uniform. Your model is not.
A remesher distributes polygons by surface area. Your object does not need them by surface area. It needs them on the silhouette and wherever the form changes direction, and it needs close to none in the middle of a flat plate.
At the 10K Meshy recommends for mobile, a hand-built low-poly spends nearly all of that budget on outline and transitions. A remesh at 10K spends it evenly, which means the curved rim that actually draws the silhouette gets the same faces per square centimetre as the flat back nobody will ever see.
The normal map is then asked to make up the difference, and it cannot. A normal map fakes surface. It has never faked an outline. A faceted silhouette stays faceted at every angle, in every engine, in every screenshot, forever.
Seams, smoothing and the cage
Three things break, and they break in this order.
Hard edges without UV seams. The rule does not bend: wherever the mesh has a smoothing split there has to be a UV seam under it, or the bake writes a gradient across the break and you get a visible band down the edge. Auto-repack cannot know which edges are meant to be hard, because the remesh had no smoothing intent to begin with. You generally get back everything smooth or everything split.
Island count. Auto-unwrap optimises for packing efficiency, not for where a seam is allowed to be seen. You get a lot of small islands, seams running across surfaces the camera reads, and at 4K with 4px padding every extra island is padding you paid for out of texel density.
The cage. This is the one that ends the argument. A remesh approximates volume; it is not projected to sit just outside the high-poly. So it floats outside the sculpt in one region and cuts inside it in another, and no single ray distance captures both. Raise the distance until the sunken parts catch and the raised parts start pulling detail off a neighbouring piece. That is the skewed, smeared normal map students send me asking what went wrong. Nothing went wrong. The low-poly was never built against that high-poly in the first place.
"Quad-dominant" is my word too
I should be fair about this, because I use the same phrase. My own base mesh listings say quad-dominant, with no ngons where they cause shading errors. That is an honest claim and it is a claim about shading. It says nothing about whether the loops follow the form.
That is what quad-dominant means everywhere, on my products and on theirs: most faces have four sides. Edge loops matter for exactly two things — deformation, and subdivision or displacement — and both of those need loops that run along the form, not loops laid out to cover the surface evenly. A solver optimising for even coverage gives you regular, beautiful, completely uninformed topology. It photographs better in a wireframe screenshot than a good low-poly does. That is worth sitting with for a second, because it is why this argument keeps getting won by the wrong side.
Where I use it anyway
I am not arguing these tools are useless. I am arguing they are sold hardest for the one job they are worst at. Four places they earn their keep:
- Decimating a scan or a sculpt for a background prop the camera never gets within a few metres of. Silhouette error nobody can see is not error.
- Mid and far LODs. Build LOD0 by hand and remesh everything under it. This is the one case where uniform density is exactly correct, because at that distance every part of the object is equally unimportant.
- Blockouts and kitbash filler, where the mesh is a placeholder for a decision and not for a bake.
- 3D printing, where topology carries no shading at all. 3D AI Studio recommends triangles rather than quads for print, which is right, and it costs 1 credit instead of 3.
And one more that is a production workflow rather than a shortcut: pulling a shape out of a generator and modelling over it, taking AI output through topology cleanup, UV rebuild, texturing and LOD creation. That is the whole subject of my Master Course AI 3D Modeling Creation and Game-Ready Workflow, and the part I would put first is choosing which outputs to keep, because most of the time this workflow saves is saved by throwing a bad result away in the first two minutes rather than the fortieth.
What would change my mind
The moment one of these ships a remesh that takes the high-poly as an input and projects to it — cage-aware, silhouette-aware, spending faces where the curvature is — the cage objection disappears and most of the rest of this goes with it. There is no reason it cannot be done. Meshy already notes that its Smart Topology only works at generation time and cannot be applied to a model that already exists, which tells you they know exactly where the hard part is.
Until then the wireframe screenshot is not the deliverable. The four mesh maps are. Bake it before you believe it.
