search

christophschuhmann / improved-aesthetic-predictor

CLIP+MLP Aesthetic Score Predictor
Apache License 2.0
854 stars 87 forks source link

Reproducing the scores in improved_aesthetics_6.5plus #12

Open hp6 opened 1 year ago

hp6 commented 1 year ago

Hi, I'm trying to reproduce the aesthetic scores presented in improved_aesthetics_6.5plus but get different results. I downloaded the improved_aesthetics_6.5plus images using img2dataset with the following command

img2dataset --url_list improved_aesthetics_6.5plus \
  --input_format "parquet" \
  --url_col "URL" --caption_col "TEXT" \
  --output_format webdataset --output_folder laion-high-resolution \
  --processes_count 16 --thread_count 64 --image_size 384 \
  --resize_only_if_bigger=True \
  --resize_mode="keep_ratio" \
  --skip_reencode=True \
  --save_additional_columns '["similarity","hash","punsafe","pwatermark","AESTHETIC_SCORE"]' --enable_wandb False

for example this image has a score of 6.8 but the simple_inference script returns 6.1. This is the env I'm using.

ChaoLi977 commented 4 months ago

I also meet the same issue. I get some images from the Visualizations of sorting all 2.37B images from LAION 5B (http://captions.christoph-schuhmann.de/aesthetic_viz_laion_sac+logos+ava1-l14-linearMSE-en-2.37B.html). I used the sac+logos+ava1-l14-linearMSE.pth to predict the scores of these images, but some scores are different with their bucket score, especially the low scores. Anyone knows the reason?