Releases

Each release has one zip per system. Do not unpack it: Blender installs the zip itself (how). Pre-releases are test versions.

Quadami v0.2.0

Published 2026-10-28 · tag v0.2.0 · on GitHub

Quadami 0.2.0: automatic quad remeshing for Blender 4.2 or newer, as an extension zip with the remesher program inside. Free software under the GNU GPL, version 3 or later.

What changed

A release with known limits, like 0.1.0, not a stable one: read Known limits below before relying on it. Your results change against 0.1.0 on most models (cleaner input, sharp edges, fewer crashes), so a model remeshed with 0.1.0 does not come out the same. Compared with 0.1.0:

• The result keeps the original's data. Under Keep from Original (a collapsed section of the panel): materials, vertex groups, colour attributes and the sharp, crease and seam marks, on by default; the UV maps, off by default. The line "Copied from the original: …" lists what came across (user guide §3).
• Messy input is cleaned up first. Pieces that only touch along a seam are welded (Blender's glTF import, text converted to mesh, an Edge Split modifier), duplicate faces are removed, points where two parts of the surface share one vertex are split, and paper-thin sliver pieces with no area are left out. These used to give cracks, thousands of empty pieces, minutes of work or an out-of-memory crash; on our test meshes they now take seconds. The Input cleanup row says what was done; your object is never changed.
• Follow Marks (next to the settings, Sharp on by default): edges you marked Sharp (or Crease or Seam, if you tick them) keep a row of quad edges along them, whatever their angle, also on smooth parts such as a design line or a sphere's equator. On our marked test meshes (Blender's cube, cylinder, cone, bevelled boxes, spheres) a quad edge follows 69 % of the marked length on average instead of 34 %, and 90-99 % on the cube and cylinder at 5,000 quads. A model that carries Sharp marks therefore comes out differently from 0.1.0; untick Follow Marks to remesh without them. The panel counts the marks, and after a run says how many marked edges it followed.
• Edges of exactly 90° (Blender's cube and cylinder) are now kept as hard edges at the default Sharp Edge of 90°. Before, rounding noise in the coordinates decided it, so a cube moved by a hair could lose most of its edges; you no longer need 89° for them.
• Fewer hangs and crashes. Some inputs made the quad extraction loop forever or crash (a mosquito model at 2,000 quads, a watch model, a lace model at 3,000 and 10,000 quads); the piece concerned now goes to the fallback extractor or fails on its own, and the rest of the model is remeshed.
• Less memory: the remesher's peak memory is lower without changing the result, by about 58 % on a model of 4,888 loose pieces, 23 % on a 2-million-triangle scan and 14 % on a 100,000-triangle sculpt at 50,000 quads.
• Missing parts say why. The warning now names the cause: parts thinner than one quad ("a limit of Target Quads, not a remeshing error"), separate pieces that got no quads, the edges of single-sided surfaces, or something else, with what raising Target Quads did on our test models.
• Quad Smoothing (Advanced, experimental): Keep Ridges slides the quads along the surface instead of moving them in 3D, so ridges, swirls and vines stay sharper. Off by default (Standard).
• The panel's text wraps to the sidebar's real width instead of being cut off.

Warnings in the panel

New or changed since 0.1.0; the others are as in 0.1.0 (user guide §6).

• Missing parts now comes with one sentence per cause: thin parts ("a limit of Target Quads, not a remeshing error"; raise Target Quads, 1.5 to 8 times as many brought the parts lost first back on 7 of 10 test models, or thicken the part), pieces with no quads, open edges of single-sided surfaces (leaves, petals, cloth), and other, which asks for a report.
• Parts thinner than one quad now says it is a limit of Target Quads at the target you asked for.
• "Followed N marked edges (Follow Marks). If you don't want quads along them, untick Follow Marks and remesh again." After every run that used marks.
• A mark on more than half of the edges is left out: on a mesh with over 1,000 edges where one mark is on more than half of them (what many FBX and OBJ imports carry), that mark is not used, with a warning in the panel and the result. Also warned: marks too dense for the quads on a part (not used there), a part remeshed again without its marks because the marked run folded, and marks that are not an edge of the mesh (ignored).
• Input cleanup lists welded seams, removed duplicate faces, split pinched points and left-out sliver pieces.
• "Could not copy data from the original: …": copying materials, groups, marks or UVs failed; the result itself is fine.

Known limits

• Thin parts (leaves, vines, cloth, blade edges) thinner than about one quad can fold, lose pieces, or lose their handles (small tunnels close up). The panel says so; raise Target Quads or thicken the part. The remesher does not put smaller quads on thin parts (tried for this release and not adopted: it either acted like a higher Target Quads for the whole model or moved the losses elsewhere). Raising Target Quads need not clear the warning: a higher target measures the parts more finely and finds thinner ones.
• Folded sculpt input is where most remaining defects on our test sculpts sit. Fold repair is off by default and leaves spots it cannot fix safely; folds wider than about 1 % of the model are not detected.
• Holes can still appear in closed sculpts, mostly next to folded input and at very high Target Quads; the Holes row shows them. Repair Folds and Re-run closed them on our folded test head, but it flattens fine detail such as swirls.
• Hard-surface models: edges still come out slightly rounded, with rows of thin quads along them at times: on Blender's cube an edge loop now follows about 90 % of the edges at 5,000 quads (72 % at 1,000), on the cylinder about 80 % of the rims (74 % at 1,000). A cone's base rim is mostly rounded off and its tip is cut off about one quad deep; with Follow Marks the cone's rim is followed at 5,000 quads but still rounded off at 1,000. A sharp spike on a sculpt is cut off the same way, and the Missing parts warning may then say the part is not thinner than a quad and ask for a report (our sculpted test box; sharp tips are planned for 0.3). The 90° change costs a little elsewhere on our test sets: a box with a sculpted relief has about 12 folded quads instead of 4 at 5,000 quads, a rotated mechanical bracket loses a handle at 1,000 quads, a vine-covered sword more folded faces at 3,000 quads.
• The cleanup can join what you meant to keep apart: open borders that lie on top of each other within about 1/100,000 of the model's size are welded (two sheets with their own vertices lying on the same side stay apart). It cannot be switched off. Splitting pinched points can add a few folded quads (a plant model: about 86-89 instead of 80-83 at 3,000 quads), and welding gave a dragon model 65-91 non-quads instead of 2 at 3,000 quads in half of our runs.
• Follow Marks: marks that cross other creases at an acute angle, or each other in many short lines, can fold the quads; such a part is remeshed again without its marks (it then takes about twice as long or more: Suzanne with marked edges at 5,000 quads took 3.6 times as long). Where marks are denser than the quads (more than a quarter of a part's planned quad edges), they are not used on that part. A diagonal loop across a cube's faces is mostly dropped at 5,000 quads. A followed rim can still step off its line at one spot, with the quad columns next to it bent towards that spot (Blender's cylinder with its top rim marked, 2,000 quads: all 32 marked edges followed, no folds or holes; Blender's cone with its base rim marked, 2,000 quads: a few rim edges step off, no folds or holes); better rims are planned for 0.3.
• Very many tiny thin pieces (a wrought-iron gate, wire ornaments) can make the remesher run for many minutes: on our iron-gate test model at 3,300 and 4,000 quads, some runs had not finished after 15 minutes, in the fallback extractor. Press Esc to cancel; Blender stays usable. Merging or deleting the smallest pieces first may help (not measured).
• A very dense lace-like piece can now come out with no quads (it is reported as an empty island) instead of the run taking many minutes or running out of memory.
• Copied data: materials, groups and marks follow the new quads, not your borders: a material border zigzags along the quads, a marked line becomes the nearest run of quad edges, and lines shorter than about one quad are dropped. UV Maps keeps each face inside one UV island, but on models with many small islands up to 6 % of the faces come out distorted. Shape keys and face sets are not kept.
• Modifiers are baked in: an enabled Mirror, Subdivision Surface or Armature is applied into the remesh and is not on the result.
• Keep Ridges quad smoothing can leave a few more quads folded over in deep grooves (a swirled cube: about 6 more out of 3,000), which is why it is not the default.
• Thousands of tiny loose parts (chain mail, scattered debris): every piece gets a minimum number of quads, so the result can have far more quads than the target, and some pieces still come out empty. Raising Target Quads does not help here; merge or delete the small parts first.
• Some very thin, complex meshes (for example the public nemetona model) can run for 30+ minutes or return far more quads than the target; cancel the run and try a higher Target Quads.
• Openings in open parts (gums, clothes) can close, open or change shape without the Holes row counting them; it counts only holes in closed parts.
• Object scale: a non-uniform scale stretches the quads; the panel warns, but apply the scale first (Ctrl+A > Scale). No mirror symmetry (remesh one half and add a Mirror modifier), guide curves or density painting yet.
• Large inputs are slow to hand over (the mesh goes through a text file both ways).
• Against 0.1.0 on our test sets (release check, corpus and closed sets, 6 runs per case on each platform): nothing crashed, and no case gets a hole or loses a piece on every run. On macOS more single runs had a hole added or closed (11 of 432 against 3, mostly the human test model at 10,000 quads, also a CAD bracket and the sword). A few validity numbers moved slightly, in different directions per platform. Folded quads went up by 5 % on Linux, 10 % on Windows and 27 % on macOS, and non-quads stay under 0.12 per 1,000 faces; on macOS almost all of the added non-quads come from one case, the vine-covered sword at 2,000 quads with the default Resolution, Quality (34-56 non-quads on 5 of 6 runs instead of none). On Windows and Linux the same sword case comes out more than 10 % under the target (all 6 Windows runs, 5 of 6 Linux runs; thin ribbons, see Thin parts).
• The same input gives slightly different results on Linux, Windows and macOS (each is repeatable on its own).
• Tried by hand on one Linux desktop in Blender 4.5, with a test build of this release's code (only these notes and the gallery changed after it): a sculpted head at 10,000 quads with both Quad Smoothing settings, an imported cube at 3,000 quads, a sword joined from several parts at 3,000 and 6,000 quads (the Sharp-marks-left-out and thin-parts warnings), Blender's default cube at 1,000 and 5,000 quads, converted text at 2,000 (seams welded, no holes), Suzanne with a material and a UV map (both copied), and a UV sphere with a Sharp equator (followed with Follow Marks on, not with it off). No crash, hang or cut-off text. Windows: see below. Linux, Windows and macOS all pass the automatic tests in Blender without a window.
• Windows: this release was not tried by hand. It passes the automatic tests (the remesher's tests, an install into a real Blender 5.2 that remeshes, the panel check and the release check), but nobody has run 0.2.0 on a Windows desktop yet; 0.1.0 was tried on one. A fix follows as a patch release if a problem shows up. The program is not signed, so SmartScreen may warn before it first runs.
• macOS is experimental and untested: nobody has run this release on a real Mac, and the program is not notarized, so macOS may block it (user guide §2). Reports from Mac users are welcome.
• Version numbers name releases only. Test builds between releases keep this version; tell them apart by the build line.

Download

Pick the zip for your system. Do not unpack it: Blender installs the zip itself (Edit > Preferences > Get Extensions, the ⌄ menu top right, Install from Disk..., then restart Blender). The user guide explains every setting and warning.

• Linux x64: any distribution with glibc 2.28 or newer (the floor of Blender 4.2 itself; e.g. Ubuntu 20.04+, Debian 10+, Fedora 29+, RHEL/Rocky/Alma 8+). Built on AlmaLinux 8, checked to need nothing newer.
• Windows x64: Windows 10 or 11. The remesher is not code-signed. Blender unpacks it itself, so Windows SmartScreen should not ask; if your antivirus quarantines autoremesher-cli.exe, restore it or build from source.
• macOS, Apple silicon (M1 or newer), macOS 11.2+: experimental, untested. It passes the automatic tests in Blender without a window, but nobody has tried it on a real Mac yet, and the remesher is not signed or notarized by Apple. Installed through Blender it should run; if macOS says it "cannot be opened because the developer cannot be verified", see the user guide, §2 (macOS). Intel Macs: build from source.

Check a download with sha256sum -c SHA256SUMS.txt (Linux), shasum -a 256 -c SHA256SUMS.txt (macOS) or Get-FileHash <zip> (Windows PowerShell), against the table above. To check that a zip was built by this project's GitHub Actions from the commit below (with the GitHub CLI): gh attestation verify <zip> -R KajetanMuczynski/quadami.

Build line. The bottom of the Quadami panel shows Add-on build v0.2.0 57c19f9 2026-10-28. Put that line in every bug report.

Source and licenses

Built by GitHub Actions from commit 57c19f9b1cdfe16ba96e1dde8cea6cd54ad14bd5 (tag v0.2.0): source, changes since v0.1.0. Each zip holds licenses/: the GPL (Quadami and libQEx), upstream AutoRemesher's MIT notice (Jeremy Hu), and the licenses of OpenMesh, oneTBB, Eigen, meshoptimizer, isotropicremesher and tinyobjloader compiled into the remesher, plus SOURCE.txt naming the source commit.

Quadami v0.1.0

Published 2026-10-27 · tag v0.1.0 · on GitHub

Quadami 0.1.0: automatic quad remeshing for Blender 4.2 or newer, as an extension zip with the remesher program inside. Free software under the GNU GPL, version 3 or later.

What changed

First public release of Quadami, the Blender add-on that began as a fork of Jeremy Hu's AutoRemesher 1.2.0, and has been substantially rewritten. It is a first release with known limits, not a stable one: read Known limits below before relying on it. Compared with AutoRemesher:

• A Blender 4.2+ extension: panel in the 3D Viewport sidebar (N, tab Quadami). The remesher runs in its own process, so Blender stays usable, Esc cancels and a crash cannot take Blender down. Modifiers are used as shown in the viewport; the original object is hidden, never changed.
• A quality report after every run: quads against the target, non-quads, holes, handles, non-manifold edges and vertices, flipped and folded faces, empty pieces, missing parts and the time, with warnings that say what to do.
• Quad extraction by libQEx; a built-in fallback extractor takes over for a piece libQEx cannot handle.
• Thin parts survive better: pieces glued along an edge (such as vines on a blade) now go through libQEx too, and two clean-up passes no longer delete strips about two quads wide.
• Repair Folds and Re-run: detects sculpt input folded over itself (twist, rotate, pinch brushes) and, only when you press it, smooths those spots before remeshing. Off by default.
• Reworked rounding and untangling of the parameterization, a map check before extraction, the output projected onto and relaxed along the input, input cleanup.
• A build line at the bottom of the panel names the exact build; it turns red when the add-on and the remesher inside it come from different builds.

Warnings in the panel

• Holes where the input was closed, non-manifold edges or vertices, empty islands: shown in red with what to try (user guide §6).
• Handles lost or added against the input (a ring's hole closing up, two surfaces joined).
• Missing parts: a part of the input (a vine, a limb, a spike) is more than three quad widths away from the result, so it was smoothed away or cut off; also islands that came out with far fewer quads than planned.
• Parts thinner than one quad: the share of the surface that is thinner than a quad, with an estimate of the target that fits a quad across about half of it, and the advice to thicken those parts. It states a risk: on our test sets most runs it fires on have folds or lost parts there, but some come out clean. Raising the target to the estimate need not clear it: a higher target measures the parts more finely and finds thinner ones.
• Folded input: N faces in K spots with the Repair Folds and Re-run button.
• Remesh to 200,000+ quads? asks first and shows the memory it needs.

Known limits

• Thin parts (leaves, vines, cloth, blade edges) thinner than about one quad can fold, lose pieces, or lose their handles (small tunnels close up). The panel now warns, but the fix is yours: raise Target Quads or thicken the part; the remesher does not yet put smaller quads on thin parts.
• Folded sculpt input is where most remaining defects on our test sculpts sit (about three quarters of them; most of the rest are on thin parts). Fold repair is off by default and leaves spots it cannot fix safely; folds wider than about 1 % of the model are not detected.
• Holes can still appear in closed sculpts, mostly next to folded input; the Holes row shows them. Very high Target Quads makes them likelier: on our folded test head, no run had holes at 10,000 quads, 2 of 6 at 20,000 and all 6 at 35,000 and 50,000 (2 to 6 holes each). Repair Folds and Re-run closed them in every run on that head, but it flattens fine detail such as swirls.
• A few more folded and flipped faces on thin parts than the 2026-10-22 test build (0e9bec8): about +5 % folded quads and +4 % flipped faces on our test sets, in the same few models on Linux, Windows and macOS. They sit on thin parts that test build deleted and this one keeps (the vines fix), mostly the sword's ribbons, the human test model and very thin closed shapes; the same change keeps more pieces and hits the quad count more often. Same advice as for thin parts above.
• Thousands of tiny loose parts (chain mail, scattered debris): every piece gets a minimum number of quads, so the result can have far more quads than the target (about 9.7x on a test model with 4,888 pieces), and some pieces still come out empty (400-429 of those 4,888). Raising Target Quads does not help here, though the panel suggests it; merge or delete the small parts first.
• Openings in open parts (gums, clothes) can close, open or change shape without the Holes row counting them; it counts only holes in closed parts.
• Sliver islands: meshes with long, almost zero-area sliver islands (for example from boolean operations or scripted cuts) can take minutes and several GB of memory; use Merge by Distance and Delete Loose first.
• The result keeps no UVs, materials, vertex groups, sharp edges, creases or seams. It keeps the source's parent, transform, collections and smooth or flat shading, and Remesh on a result remeshes the hidden original (the panel's Source field).
• Object scale: a non-uniform scale stretches the quads; the panel warns, but apply the scale first (Ctrl+A > Scale).
• Hard-surface models: sharp edges come out rounded. On Blender's default primitives an edge loop follows about half to two thirds of a box or cylinder edge from 1,000 quads up (less at lower counts); a cone's base rim is mostly rounded off and its tip is cut off about one quad deep. Edges marked Sharp in Blender are not used. Edges of exactly 90° (the cube, the cylinder) sit on the default threshold: Sharp Edge 89° keeps more of them, but on detailed parts lower angles make more triangles (user guide §7). No mirror symmetry (remesh one half and add a Mirror modifier), guide curves or density painting yet.
• Large inputs are slow to hand over (the mesh goes through a text file both ways).
• Tried by hand with this release's build on one Linux desktop: a sculpted head, a cube, and a sword joined from ten parts at 3,000 to 35,000 quads, also with Repair Folds; on one Windows 10 desktop: the cube at 3,000 quads. What the result keeps (parent, shading, the Source field, the scale warning) is covered only by the automatic tests. Linux, Windows and macOS all pass the automatic tests in Blender without a window.
• Windows: the program is not signed, so SmartScreen may warn before it first runs.
• macOS is experimental and untested: nobody has run this release on a real Mac, and the program is not notarized, so macOS may block it (user guide §2). Reports from Mac users are welcome.
• Version numbers name releases only. Test builds between releases keep this version; tell them apart by the build line.

Download

Pick the zip for your system. Do not unpack it: Blender installs the zip itself (Edit > Preferences > Get Extensions, the ⌄ menu top right, Install from Disk..., then restart Blender). The user guide explains every setting and warning.

• Linux x64: any distribution with glibc 2.28 or newer (the floor of Blender 4.2 itself; e.g. Ubuntu 20.04+, Debian 10+, Fedora 29+, RHEL/Rocky/Alma 8+). Built on AlmaLinux 8, checked to need nothing newer.
• Windows x64: Windows 10 or 11. The remesher is not code-signed. Blender unpacks it itself, so Windows SmartScreen should not ask; if your antivirus quarantines autoremesher-cli.exe, restore it or build from source.
• macOS, Apple silicon (M1 or newer), macOS 11.2+: experimental, untested. It passes the automatic tests in Blender without a window, but nobody has tried it on a real Mac yet, and the remesher is not signed or notarized by Apple. Installed through Blender it should run; if macOS says it "cannot be opened because the developer cannot be verified", see the user guide, §2 (macOS). Intel Macs: build from source.

Check a download with sha256sum -c SHA256SUMS.txt (Linux), shasum -a 256 -c SHA256SUMS.txt (macOS) or Get-FileHash <zip> (Windows PowerShell), against the table above. To check that a zip was built by this project's GitHub Actions from the commit below (with the GitHub CLI): gh attestation verify <zip> -R KajetanMuczynski/quadami.

Build line. The bottom of the Quadami panel shows Add-on build v0.1.0 2a0ef4b 2026-10-27. Put that line in every bug report.

Source and licenses

Built by GitHub Actions from commit 2a0ef4be40c944a0fac96716d608821c4eee1c55 (tag v0.1.0): source, changes since 4a49f82. Each zip holds licenses/: the GPL (Quadami and libQEx), upstream AutoRemesher's MIT notice (Jeremy Hu), and the licenses of OpenMesh, oneTBB, Eigen, meshoptimizer, isotropicremesher and tinyobjloader compiled into the remesher, plus SOURCE.txt naming the source commit.