I noted that you used classifier-free guidance for RealEstate10K dataset. Do you have any ablation study on this design choice? How much does this contribute to the final performance? This is actually pretty expensive as it requires another volume rendering step for the unconditional generation part, so I am wondering if it's essential to get DFM to work for scenes or it is more about improving over a working baseline.
hi,
I noted that you used classifier-free guidance for RealEstate10K dataset. Do you have any ablation study on this design choice? How much does this contribute to the final performance? This is actually pretty expensive as it requires another volume rendering step for the unconditional generation part, so I am wondering if it's essential to get DFM to work for scenes or it is more about improving over a working baseline.
Best shengyu