Marigold depth model crash – SciPy import error due to embedded Python

[xmoto4]

Hi,
When using the Marigold depth model, VisionDepth3D successfully extracts the video into frames, but it crashes every time depth inference begins. According to the crash log, Marigold attempts to run in ensembling mode and tries to import SciPy:

ImportError: Make sure to install scipy if you want to use ensembling.

SciPy is installed in the system Python, but it appears that the VisionDepth3D embedded Python environment does not see or use the system Python installation. The settings.json and AppData folders do not contain a marigold_ensembling setting, so it appears that ensembling is hard-coded and cannot be disabled.

The crash log:

=== 20260116-170448 === Traceback (most recent call last): File "threading.py", line 1043, in _bootstrap_inner File "threading.py", line 994, in run File "core\render_depth.py", line 2302, in process_video2 File "core\render_depth.py", line 1838, in process_images_in_folder File "core\render_depth.py", line 463, in _run_pipe_or_tile File "core\render_depth.py", line 1151, in diffusion_pipe File "torch\utils\_contextlib.py", line 120, in decorate_context return func(*args, **kwargs) File "diffusers\pipelines\marigold\pipeline_marigold_depth.py", line 448, in __call__ File "diffusers\pipelines\marigold\pipeline_marigold_depth.py", line 229, in check_inputs ImportError: Make sure to install scipy if you want to use ensembling.

Could you please clarify:

• Is Marigold ensembling currently hard‑coded?
• Do you plan to include SciPy in the embedded Python environment?
• Will a future update allow Marigold to run in single‑pass mode without SciPy?

Thank you for your work — VisionDepth3D is a fantastic tool, and I would love to use Marigold reliably as well.
Best regards,
Tamás

[xmoto4]

Hi again,
I wanted to provide an update regarding the Marigold depth model issue I previously reported.
It turns out that the crash only occurred when using the Windows installer version of VisionDepth3D. After switching to a full manual installation (Python 3.12, CUDA 12.8, cuDNN 9.x, and the CUDA‑enabled PyTorch build), Marigold now runs perfectly — including depth inference — and the SciPy‑related error no longer appears.
So the problem seems to be specific to the embedded Python environment bundled with the Windows installer, not to VisionDepth3D or Marigold itself.
Everything is working beautifully now.
Thank you for your work on this project — it’s an excellent tool, and I’m glad to be able to use Marigold properly.
Best regards,
Tamás

[VisionDepth]

Thanks for finding this bug, I believe when I made the exe SciPy wasn't properly packaged in the dependencies when compiling the exe, someone in the subreddit has brought this to my attention, thanks again and will work on the patch when I have the time