It's not a issue, it's to help people reproduce it

[qianx77]

I found that although this is a single step diffusion, the memory consumption during training is still quite large. Considering that I only have RTX3090, I will record some of my settings, hoping to help those in need.

1、--train_batch_size need to set 1.
2、--rank_lora need to set from 64 to 16

nohup accelerate launch  --config_file config/config.yaml  --gpu_ids 0,1,2,3 --num_processes 4 --main_process_port 57079 --mixed_precision="fp16" train/train.py \
  --pretrained_model_name_or_path=$MODEL_NAME  \
  --teacher_lora_path=$TEACHER_MODEL_NAME \
  --train_batch_size=1 --rank=64 --rank_vae=64 --rank_lora=16  \
  --num_train_epochs=200 --checkpointing_steps=5000 --validation_steps=500  --max_train_steps=200000 \
  --learning_rate=5e-06  --learning_rate_reg=1e-06 --lr_scheduler="cosine_with_restarts" --lr_warmup_steps=3000 \
  --seed=43 --use_default_prompt --use_teacher_lora --use_random_bias \
  --output_dir=$OUTPUT_DIR \
  --report_to="wandb" --log_code --log_name=$LOG_NAME \
  --gradient_accumulation_steps=1 \
  --resume_from_checkpoint="latest" \
  --guidance_scale=7.5  > $OUTPUT_LOG 2>&1 & \

[XuBao12]

nice advice

[spongiaaa]

Can such a setting reproduce results close to those reported in the article?

[qianx77]

Can such a setting reproduce results close to those reported in the article?

I try to reproduce it using my own data, but the result is bad. I don't change anything except the train_batch_size and rank_lora . I don't know why.