For I23D, what are the correct output of stage1 and input for stage2?

[KaqiKaqi]

After running the inference script of stage1 of I23D, I got the following output from stage1:

├── stage-1 │   └── dino_img │   ├── args.json │   ├── blue_cat-input │   │   ├── imgcond.jpg │   │   ├── sample-0-0.glb │   │   └── sample-0-0.ply │   ├── cake-input │   │   ├── imgcond.jpg │   │   ├── sample-0-0.glb │   │   └── sample-0-0.ply

Then when running the inference of stage2, it said there is no sample-0-1.ply file in the corresponding directory. Did I miss anything in stage1 inference? Thanks a lot.

[NIRVANALAN]

The output structure is fine, it looks like the default batch size issue. The k sample-0-{k}.ply indicates the index of the batch here. Setting batch size to larger in stage-1 shall fix this issue. I will fix the code later accordingly.

[KaqiKaqi]

The output structure is fine, it looks like the default batch size issue. The k sample-0-{k}.ply indicates the index of the batch here. Setting batch size to larger in stage-1 shall fix this issue. I will fix the code later accordingly.

Thanks for your explanation.

I modified the i23d-stage1.sh: set batch_size=2 on line 16 and --eval_batch_size 2 on line 164. But sample-0-1.ply was still not generated after stage-1 inference. The output directory of stage1 inference is same for each subdir(e.g. sword/), but some images do not have correspondin output sub directory.

Besides, I find https://github.com/NIRVANALAN/GaussianAnything/blob/f720d9882b5b872dab7b4e0eb7f708f21c7b44ae/nsr/lsgm/flow_matching_trainer.py#L2118

this range(2) in the code, after which sample-0-1.ply 's absolute path is loaded by trimesh.load().