Two releases landed about two hours apart on August 12, 2026: WebODM 3.2.8 and ODM 3.6.2. If you follow both projects casually, it’s easy to assume that’s one coordinated release across the stack. It isn’t. These are two different forks — the split I covered back in June — shipping independently, on their own schedules, to their own audiences. Here’s what actually changed in each, verified directly against the GitHub release records and the underlying pull requests, and what it means if you’re running either one.
WebODM 3.2.8: A Four-Year-Old Request Finally Ships
WebODM 3.2.8 (github.com/WebODM/WebODM, published 2026-08-12 19:14 UTC) is the actively-developed fork’s UI release. Three changes:
Permanent overlays with DXF export (PR #1998) is the headline feature, and it has a longer backstory than the changelog entry suggests. Before 3.2.8, any vector layer you added to a project — site boundaries, flight-line references, annotation markup — was temporary: close the tab and it was gone. Users had been asking for persistent overlays since at least January 2022, when maintainer Piero Toffanin confirmed on the community forum that “all add-on layers are temporary” with no plan in place to change it. The formal tracking issue (#1946) wasn’t filed until May 23, 2026 — by Toffanin himself, one line: “Could be useful in certain scenarios to be able to add permanent vector layers.” The PR closed it two and a half months later. Overlays are now persistent project data, and DXF joins GeoJSON and Potree as an export format.
For anyone delivering to a civil engineer or CAD-based client, DXF is the format AutoCAD and Civil 3D actually expect. Previously, getting a WebODM vector overlay into a CAD workflow meant exporting GeoJSON and round-tripping it through QGIS for format conversion. Now it’s a direct export. It’s not a contour or breakline extraction tool — you still need dedicated survey software for that — but for boundary and annotation layers, one conversion step disappeared.
A quotas command (PR #2000) was added for querying usage-limit information from the command line. The PR description doesn’t say what consumes it, and no changelog or forum post connects it to a specific feature — treat it as infrastructure for account/usage tracking rather than something that changes your day-to-day workflow.
A toaster-notification link display fix (PR #2001) — cosmetic, not worth planning an upgrade around on its own.
| WebODM 3.2.8 change | PR | Why it matters |
|---|---|---|
| Permanent overlays + DXF export | #1998 | 4-year-old request closed; direct CAD export path |
quotas CLI command | #2000 | Usage-limit querying (purpose not documented) |
| Toaster link display | #2001 | Minor UI polish |
If you’re running the standard WebODM installer or Lightning, you’re on this fork by default — see the ODM vs ODX naming guide if that sentence doesn’t make sense yet.
ODM 3.6.2: Five Fixes, One of Them a Silent Data-Loss Bug
ODM 3.6.2 (github.com/OpenDroneMap/ODM, published 2026-08-12 21:06 UTC) is the OpenDroneMap nonprofit’s standalone processing engine — the one you’re running if you pulled opendronemap/odm directly, or if NodeODM is talking to it without a WebODM UI in front. The release notes are blunt about why it exists: “to fix some high priority issues found in the 3.6.1 release,” crediting three community members for reporting them.
--build-overviews crashed on georeferenced runs (PR #2059, then #2061). 3.6.1 introduced a config string built with a missing space, so gdaladdo received a malformed argument. The first fix patched the string — but that exposed a deeper problem: 3.6.1 had also set overview color encoding to PHOTOMETRIC=YCBCR, which GDAL requires to be exactly 3 bands. ODM orthophotos always carry a 4th alpha band, so this was never going to work on any georeferenced dataset. PR #2061 reverted the YCBCR change entirely rather than patch around it.
GPU builds silently fell back to CPU (PR #2060). The Ubuntu 24.04 migration in 3.6.1 dropped a CMake DEPENDS line, which let the build system delete the compiled PyPopsift GPU module right after building it. The result: GPU Docker images kept running, just on slower CPU-based SIFT feature matching, with no error to indicate anything was wrong. Two independent community forum threads (#26715, #26711) reported the symptom before this fix shipped.
DJI video telemetry could vanish entirely (PR #2065, plus a related fix for issue #2063). The SRT subtitle-track parser assumed integer shutter-speed values; fed a float, it threw an exception — and a separate logging bug in the same code path swallowed that exception instead of surfacing it, silently dropping all GPS telemetry for the video. A second, unrelated bug in the same parser wrote pre-GPS-lock frames as coordinate (0,0) instead of null, which can quietly corrupt a bundle adjustment rather than just omitting a point. Both are fixed in 3.6.2.
Multispectral TIFF writes could fail outright (PR #2066). Under NumPy 2, a float64 irradiance scalar silently promoted float32 reflectance data to float64 during estimation, and GDAL then rejected the resulting TIFF for radiometric datasets. This is a hard failure, not a quality regression — if you hit it on 3.6.1, the run didn’t quietly produce worse output, it didn’t produce output at all.
| ODM 3.6.2 fix | What actually happened on 3.6.1 |
|---|---|
--build-overviews crash + YCBCR revert (#2059, #2061) | Overview generation failed on every georeferenced run |
| GPU SIFT fallback (#2060) | GPU builds silently ran on CPU, no error shown |
| SRT telemetry parser (#2065, #2063) | Could silently drop all GPS data from DJI video captures |
| float32 reflectance (#2066) | Multispectral TIFF writes could fail outright |
Two of the five PRs disclose AI-assisted diagnosis, and the linked issue that spawned a third (#2063) explicitly credits an AI-assisted review pass — Claude Code — for finding the bug in the first place. Not something every release discloses, and a sign of where open-source bug-hunting is headed.
If you’re still on 3.5.6, skip straight to 3.6.2. 3.6.0 never produced installable artifacts — no Windows installer, no Docker images. 3.6.1 was the first installable release since 3.5.6 and shipped a real fix worth having — a scale-corruption bug in --align traced to a feature-type default change (sift to dspsift) that had quietly degraded alignment accuracy from 0.999 down to 0.033 on reference datasets — but it arrived bundled with the five regressions above. 3.6.2 is the first release in that line that’s both installable and clean, while carrying forward the Ubuntu 24.04, Python 3.12, and CUDA/GDAL 3.11.1 upgrades.
Why a Same-Day Double Release Is Easy to Miss
Both are public GitHub releases, timestamped and documented — neither is hidden. But neither generated a dedicated announcement thread on the community forum or a blog post beyond the release page itself, so nothing pushes them in front of you. These are routine point releases, not marketing moments, and that alone is enough to miss them if you’re not checking GitHub directly. The bigger issue is structural: WebODM and ODM stopped being one project in April 2026, so “check for ODM news” now means checking two separate GitHub organizations with two separate release cadences — and it’s easy to check one and assume it covers both. It doesn’t.
Bottom line: WebODM 3.2.8 is worth the upgrade if you export overlays for CAD-adjacent deliverables — DXF export removes a real round-trip step, on a feature people have wanted since 2022. ODM 3.6.2 isn’t optional if you’re anywhere near 3.6.1 — it fixes a bug that can silently drop DJI video telemetry and one that can outright fail multispectral processing, neither of which announces itself with a clear error. Check which fork you’re actually running before you decide which changelog applies to you.