Outputs are identical chair shapes regardless of input shape

[maa989]

Hello, first and foremost thanks so much for sharing your work — it’s a really impressive project and the paper was a great read. I’ve been experimenting with it on some of my own meshes and wanted to double-check whether I might be missing something.

I followed the setup instructions, downloaded the provided checkpoints, and used my own chair models as input. Each OBJ was normalized as instructed.

Here’s the command I ran:

python geocode/geocode.py test \
  --blender-exe ~/Blender/blender-4.2.3-linux-x64/blender \
  --blend-file ./blends/procedural_chair.blend \
  --models-dir ./models \
  --dataset-dir ./fresh \
  --input-type pc \
  --phase test \
  --exp-name exp_geocode_chair \
  --normalize-pc

Everything runs smoothly and the model loads procedural_chair_epoch585.ckpt without errors, but the resulting output always looks like the same base chair shape — only the overall dimensions seem to change, not the structure.

Is this expected (e.g., due to the procedural program’s limited variability), or could I be missing a step that enables more diverse reconstructions? Here is a screenshot of some examples (input objs on left hand side of each pair):

Thanks again for making this work public — it’s been great to explore and learn from.

[ofekp]

Thank you for taking interest in our project.
I want to see if this is a prediction issue or a Blender geo-nodes issue.
Can you please verify that the YAML files that are generated from the prediction stage are changing only in the dimension of the shape?