Image-editing performance inconsistent the paper

[peacewang017]

I tried the interleave generation task (image editing) by directly running the unchanged repo under MMaDA-Parallel-A/:

python inference.py 
--checkpoint tyfeld/MMaDA-Parallel-A     
--vae_ckpt tyfeld/MMaDA-Parallel-A    
--prompt "Replace beer with a cup of coffee and make the keyboard space gray"     
--image_path examples/image.png    
--height 512     
--width 512    
--timesteps 128     
--text_steps 256     
--text_gen_length 256     
--text_block_length 32    
--cfg_scale 0     
--cfg_img 4.0     
--temperature 1.0     
--text_temperature 0     
--seed 42 
--output_dir output/results_interleave

According to the paper, I should get a edited image faithful to the original image, like this:

But it turns out I get this:

Is there something wrong with my way of reproducing it?