search

omer11a / bounded-attention

MIT License
56 stars 7 forks source link

get baseline results #5

Closed dain5832 closed 4 days ago

dain5832 commented 2 months ago

Hi, thanks for the wonderful work!

I would like to ask how can I get the baseline results using the code in the repo.

I tried to set init_step_size and final_step_size to 0, so that the step size of latent updates process becomes 0. However, the result seems to include the box condition. 286_0_0_boxes

omer11a commented 2 months ago

To recreate the baseline results, please use the hyper-parameters in our example in run_xl.py, i.e. init_step_size=25 and final_step_size=10. Setting them to zero cancels the optimization phase, which is necessary for proper alignment with the prompt and layout. The script creates two images, one that includes the boxes and one without them.