Repository navigation
Releases
Quadami v0.2.1
Quadami 0.2.1: 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 fix release on top of 0.2.0, with one new button and a stricter fold count. Like 0.2.0 it is a release with known limits,
not a stable one. On all 770 runs of our standard test sets the result is the same as with 0.2.0,
quad for quad; only models of the kind in the first point below can come out differently. Compared with 0.2.0:
- Models with very many tiny thin pieces no longer run for many minutes. On a wrought-iron gate
or a lace-like model, the fallback quad extraction could trace tens of thousands of quads for a
piece meant to get about a hundred, which took many minutes or ran out of memory. Such a piece is
now left out instead (it shows as a piece with no quads), and a later attempt at a higher count
usually fills it. Our iron-gate test model now finishes in 9-20 seconds at 3,300 and 4,000 quads
(12 of 12 runs). - Copy Report (bottom of the result box): puts one block of text on the clipboard for a bug
report: the build line, Blender and system versions, the run's settings, the input's size, every
row and warning of the result box, the time of each step, the remesher's last output and the full
report. It holds no mesh data, and your user name is removed from folder paths. It also works after
a failed run. Paste it into the bug report form or an e-mail (user guide §6). - Folded faces now also counts twisted quads: quads whose two halves face apart along either
diagonal, which the old corner test missed. The row readsN (T twisted). Quads bent along a
sharp edge of your model (90° or more) are not counted. The mesh is the same as before; only the
number is higher on some models (on our test sets, about 1.4 instead of 1.1 per 1,000 quads). - Imports marked Sharp all over no longer leave zigzag lines on the result. 0.2.0 already left
such marks out of Follow Marks, but Keep from Original still copied them, so an FBX or OBJ
import that marks most edges Sharp got zigzag sharp paths all over the result. A mark on more than
half of the edges of a mesh with over 1,000 edges is now not copied either.
Warnings in the panel
-
Folded faces includes twisted quads and shows how many of them there are (
N (T twisted)), so
it can be higher than 0.2.0 showed for the same result. -
A mark on more than half of the edges is left out: the warning now also says that the mark is
not copied onto the result.
The other warnings are as in 0.2.0 (user guide §6).
Known limits
As in 0.2.0 (thin parts, folded
sculpt input, holes, hard-surface edges, the cleanup, Follow Marks, copied data, modifiers baked in,
Keep Ridges), except for the long runs on very many tiny thin pieces, fixed above. In their place:
- Very many tiny thin pieces (a wrought-iron gate, wire ornaments, lace) can come out with some
pieces missing (red Missing parts). Raise Target Quads or delete the smallest pieces. If a run
still takes long, press Esc to cancel; Blender stays usable. - Copy Report keeps the remesher's output only for the last run since Blender started. Read the
text before you send it: folder paths outside your home folder are kept as they are. - Against 0.2.0 on our test sets (release check, corpus and closed sets, 6 runs per case on each
platform, run 41224250346 on2b858a3): passed on Linux, Windows and macOS. No structural check
fails on more cases than with 0.2.0, and no validity number is worse. Changes merged after that
commit leave the output mesh byte-identical on 770 Linux runs of our test sets, so the check was not
repeated. - Tried by hand on Linux (Ubuntu 24.04, Blender 4.5, test builds of this release's code): Blender's cube at 2,000 quads with Copy Report (no user name in the
text), a sword joined from several parts with Sharp on most edges at 3,000 quads (marks left out
and not copied, no zigzag sharp lines), and Copy Report after a run cancelled with Esc. No crash,
hang or cut-off text. After a cancelled run, Copy Report shows "0 islands" for the input (not counted
yet), not the real number. Checked again on the final build: the sword at 10,000
quads shows Folded faces 39 (24 twisted), matching the report. - Tried by hand on Windows (a real Windows desktop, this release's CI build; build line correct,
Defender silent, 0 detections on VirusTotal for the download zip): Blender's monkey at 5,000 quads,
Fast 5,005 quads (1 folded, 0.7 s) and Quality 5,093 quads (0 folded, 1.8 s); a head at 10,000
quads, 9,854 quads (4 folded, 3 twisted); the cube at 3,000 quads, 2,948 quads (1 folded,
1 twisted); the sword at 3,000 quads, 2,861 quads (44 folded, 24 twisted). Every result all quads,
no holes or non-manifold edges. The 300,000-quad warning appears and cancels, Esc cancels a run,
Ctrl+Z undoes it, and no files are left in %TEMP%. It also passes the automatic tests (the remesher's tests, an
install into a real Blender 5.2 that remeshes and the panel check). The program is not code-signed
(see Download). - 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.
| System | File | SHA-256 |
|---|---|---|
| linux-x64 | quadami-0.2.1-linux-x64.zip |
2f094d973c1c6bada7155c2b7ffbe77219dfdeaabbadd4ffb3a0e30ccd2b48af |
| windows-x64 | quadami-0.2.1-windows-x64.zip |
336469c0b5bf66e5f7b9cf714fbaeb01879b3119726b9d29cd98f1d20bee535c |
| macos-arm64 | quadami-0.2.1-macos-arm64.zip |
757b2ff4c99e40cb70029b44edc5ddf10d6434cf04cf37c0a10051a0d6214501 |
- 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 quarantinesautoremesher-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 v0.2.1 09b2e20 2026-11-12. Put that line in
every bug report.
Source and licenses
Built by GitHub Actions from commit 09b2e20d53fb86e51ee24624bd13a87729ec2885 (tag v0.2.1):
source,
changes since v0.2.0.
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 v0.2.0
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 s...
Quadami v0.1.0
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.
| System | File | SHA-256 |
|---|---|---|
| linux-x64 | quadami-0.1.0-linux-x64.zip |
10fecd1c9cb3627c5f01af52c1a4b0f7a66f64c4a8c50230ccb22b894f47c873 |
| windows-x64 | quadami-0.1.0-windows-x64.zip |
f49d254a19decaadef767abfe04aa90efb39bf8fb7a7522ae461e6f4d2747234 |
| macos-arm64 | quadami-0.1.0-macos-arm64.zip |
c0462512cfd43dc9c054620218bed7cdc56d73ccaa2293e66b0b4b9b01e3ae21 |
- 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 quarantinesautoremesher-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.