Some results for decoder on CUB

[mhh0318]

I've trained a small model for the DALLE.2 Decoder part on CUB dataset.

output size is 64*64.

The sample process is the p_sample instead of DDIM. I would update the DDIM sample result later.

The major codes is from GLIDE and some of them are inspired by this repo.

Given Image for CLIP embedding

Generated Sample

Training Loss

[xinmiaolin]

Hi,

Thanks for sharing the information. Is it possible to know the following:

1- what preprocessing did you apply on the cub 200 dataset?
2- did you use image embeddings to the decoder?
3- the unet configurations in the decoder?

Thank you very much!