Generated, decoded structures all seem to be of P(1) symmetry about xtal2png HOT 4 OPEN

materialsproject/pymatgen#1764 mentioned same issue.

Looking a bit closer at the structures, probably has to do with the presence of distances less than 0.2 between differing sites, causing atoms to "collide".

structures[0].distance_matrix

array([[0.        , 0.97919993, 1.18065832, 1.05236272, 1.2924594 ],
       [0.97919993, 0.        , 1.94263668, 1.85350771, 2.08496016],
       [1.18065832, 1.94263668, 0.        , 0.17392385, 0.15348728],
       [1.05236272, 1.85350771, 0.17392385, 0.        , 0.25348576],
       [1.2924594 , 2.08496016, 0.15348728, 0.25348576, 0.        ]])

Note that these small distances are localized to $H$ and $He$.

structures[0].atomic_numbers

(47, 24, 1, 1, 2)

from xtal2png.

Looping through all 16 generated samples

Interestingly with the higher symprec of 0.1 instead of 0.01 more None-s are produced

space_group = []
for s in structures:
    try:
        space_group.append(s.get_space_group_info(symprec=0.1)[1])
    except Exception as e:
        space_group.append(None)
        pass
print(space_group)

[1, None, None, None, 1, 1, None, None, 1, None, None, 1, None, 1, 1, None]

from xtal2png.

Better now, but still not great:

[1, 1, 1, 1, 1, 1, 8, 1, 1, 1, 1, 1, 1, 1, 1, 1]

with relax_on_decode, decode_as_primitive, and 127k iterations.

The formulas:

which seem somewhat better.

The images (sample-127.png):

from xtal2png.

from Kevin:

yes, also in your case, if you'd encode the space group in the image, you could write components for the loss that penalize mismatching angles. In other contexts, I have seen such extra loss terms to help a lot

from xtal2png.