your pixels are not broken - it’s just that the box size is set according to the desired final resolution to use for ab initio (default 12 Å). This speeds computation. In your case, that means a box of 32px.
If you want to get the same reconstruction in your original box, you could take the output of ab initio and perform a homogeneous reconstruction job.
Alternatively, you may want to play with the ab initio parameters anyway - I often find that for small proteins in nanodisc or detergent, starting at higher resolution works better. Often starting at 9Å and going to 7 or even 5 Å final works well.
It also looks like you have enforced symmetry (C2?) during ab initio. This can be useful in certain cases, but I wouldn’t recommend it for a first attempt.