v1.0.1

Both default depth models changed, and the interface stopped hiding things
from you. If you are on v1.0.0, this is worth taking.

Upgrading: download the installer and run it again. No uninstall first. It
recognises the existing install, keeps Python, ffmpeg and PyTorch, and replaces
the app. Your settings and downloaded models are untouched. On v1.0.0 that was
impossible — running the installer twice crashed on the first file it tried to
unpack.

New depth models

The old default could not resolve thin structures and made flat walls wobble.
Measured on one 360 photo, six cube faces, everything else identical:

	chair gap (→1.0)	wall wobble (→0%)	floor rms (→0%)	thin trim
V2 Base (was the default)	1.76	103.2	40.7	lost
Depth Anything V3 small — now the video default	1.42	20.2	28.1	lost
Depth Pro — now the still-image default	1.49	53.1	30.1	kept

Two defaults because the right trade differs: video wants a model small and
fast enough to run thousands of times, a still wants the best single frame.
Judged in a headset the order was Depth Pro, then V3, then V2 — the opposite of
what the geometry numbers say, because none of them measures thin structure.

Depth Pro is licensed apple-amlr, which is not an open-source licence.
It is the stills default because it measured best, not because it is the most
freely licensed option. --depth-backend depth-anything-v3 uses the Apache-2.0
model for stills too. It also downloads 3.6 GB on first use, announced before
the wait rather than after.

Photo tiling now defaults to 1 rather than 3. Tiling helped V2 and measurably
hurts both of these — Depth Pro's wall wobble more than triples under it.

The GPU

On Windows, ONNX now runs through DirectML whatever GPU you have, including
NVIDIA. onnxruntime-gpu looks correct on an NVIDIA card and is not: it ships
no sm_120 kernels, so on any RTX 50 series card it fails with no kernel image
is available for execution on the device. Measured on an RTX 5070 Ti: 1.91 s →
0.15 s for a whole frame, with bit-identical output.

The installer verifies this by running a graph on the GPU rather than trusting
the provider list.

Fixes

• The preview panel stayed empty after converting a photo. The file was
  written correctly; Qt was refusing to decode it, because a full-size stereo
  JPEG needs more memory than its limit allows. It displays now, and a decode
  that does fail says so instead of showing nothing.
• --output-width was ignored for stills. The Resolution dropdown was live
  on a photo job and did nothing — you chose 5760 and got the full size back.
• The window looked frozen while a model downloaded. It now says what it is
  doing before the first frame exists.
• The installer cached the wrong model — 400 MB of a model that stopped
  being a default, while the one that is left to download on first use.
• --version exists. Note that v1.0.0 reports itself as 0.1.0: the
  constant was not kept in step with the tag until now. From v1.0.1 they match.

Initial release

stereo360 v1.0.0

Convert monoscopic 360 video and photos into stereoscopic 3D for VR
headsets, as full 360 or VR180.

Windows install is one file: download 'Install stereo360.bat' below and
double-click it. It installs a private Python, the right PyTorch for
your GPU, the interface, ffmpeg and the depth model, with nothing added
to PATH and no administrator rights. It appears in Settings > Apps and
uninstalls from there.

Three accelerator paths, each exercised on real hardware: CUDA with the
build checked against your card's architecture, DirectML for AMD and
Intel via an ONNX model built during installation, and CPU.