Ti does not seem to be working correctly

[aiXander]

Hey @Littleor I've been trying your training script here but am not getting great results so far.. Are you able to successfully train things into a token?

I've been running with the following command:

accelerate launch --config_file "accelerate_config/deepspeed_zero3_offload_config.yaml"  textual_inversion_flux.py \
  --pretrained_model_name_or_path="FLUX.1-dev" \
  --train_data_dir=$DATA_DIR \
  --learnable_property="object" \
  --initializer_token="person" \
  --placeholder_token="<person>" \
  --validation_prompt="a photo of <person> climbing mount Everest" \
  --num_validation_images=2 \
  --validation_steps=200 \
  --save_steps=200 \
  --mixed_precision="bf16" \
  --resolution=512 \
  --train_batch_size=1 \
  --gradient_accumulation_steps=2 \
  --use_8bit_adam \
  --max_train_steps=800 \
  --learning_rate=5.0e-04 \
  --scale_lr \
  --num_vectors=2 \
  --lr_scheduler="constant" \
  --lr_warmup_steps=0 \
  --train_text_encoder_2 \
  --output_dir="./person_test_v4" \
  --center_crop \
  --gradient_checkpointing \
  --cache_latent

[Littleor]

After conducting some general case checks, it is indeed confirmed that there is an issue. I suspect it may be due to a problem with the T5 Encoder. I will investigate this issue further and fix it as soon as possible.