Trained model generates noisy images

[jh27kim]

Hi, thank you for releasing the work and codebase. It is indeed a good place to start new researches.

I tried to reproduce the reported results, but denoising does not seem to work...
The image is obtained at epoch 130.

Here' the command I used for training. I would appreciate any advice!

torchrun --nproc_per_node=8 --nnodes=1 --node_rank=${NODE_RANK} --master_addr=${MASTER_ADDR} --master_port=${MASTER_PORT}  main_mar.py 
--img_size 256 --vae_path pretrained_models/vae/kl16.ckpt --vae_embed_dim 16 --vae_stride 16 --patch_size 1 \
--model mar_large --diffloss_d 6 --diffloss_w 1024 
--epochs 800 --warmup_epochs 100 --batch_size 64 --blr 1.0e-4 --diffusion_batch_mul 4 
--output_dir ${OUTPUT_DIR} --use_cached --cached_path {$CACHED_PATH}