unable to reproduce the Dreamgen Bench metrics in the paper

[cbtogu]

Hello, and thank you for your work. However, I was unable to reproduce the Dreamgen Bench metrics reported in the paper using the Cosmos-Predict2-14B-Sample-GR00T-Dreams-GR1 model. Did I miss something here? Thank you.
Here is the procedure I followed:

1. Generated videos by running inference with Cosmos-Predict2-14B-Sample-GR00T-Dreams-GR1;

python -m examples.video2world_gr00t \ 
--model_size 14B \ 
--gr00t_variant gr1 \ 
--batch_input_json dream_gen_benchmark/gr1_object/batch_input.json \ 
--disable_guardrail

2. Evaluated them on the Dreamgen benchmark with zero_shot set to false.

python -m dreamgenbench.eval_sr_qwen_whole \
    --video_dir "$video_dir" \ 
    --output_csv "$csv_path" \  
    --device "$device" \ 
    --zeroshot false

The results I obtained are shown below,

alongside those reported in the paper:

[suhyeok-jang]

same question.