Why train for quantization AFTER image representation?

[ChaosAdmStudent]

At present, the way the README states this, the quantization training script (train_quantize) is executed after train.py has already generated the gaussian model for a given image. We then run the quantization code after we have this model_path. However, train_quantize does quantize-aware training again, which involves the same gaussian projection/rasterization operations. This essentially seems like double work.

Is there a particular problem if we just run the quantized training script for gaussians starting from scratch (without stored gaussian model)?

My main problem here is that you end up training for 2*N iterations: N iters for image representation and then N for quantization-aware training. This will considerably increase encoding time.