[One-page summary] NerfDiff: Single image View Synthesis with NeRF guided Distillation from 3D aware Diffusion by Gu et al.

● Summary: Use diffusion model to generate multiple view images for NeRF

 

● Approach highlight

• Training phase: Train a diffusion model to generate images corresponding to multiple views of an object in the training phase.
• Finetuning phase: The diffusion model learned in the training phase is used to train the NeRF by creating multiple views of the image

 

● Main Results

● Discussion

• In my opinion, if the objects learned in the training phase and the objects in the finetuning phase are different, sample generation will not work well.
• Require multiple views of the object at training time
• Heavy model(NeRF + Diffusion)