Skip to content

Latest commit

 

History

History
54 lines (53 loc) · 2.65 KB

CelebASingle.md

File metadata and controls

54 lines (53 loc) · 2.65 KB
Raw

CelebA (single attributes)

This experiment compares the diversity and image quality of vanilla Stable Diffusion, ITI-GEN, HPS and HPS with negative prompting.

Training

Train ITI-GEN tokens

You need to run jobfiles/celeba_single/iti_gen/train/${attribute}.sh, for example jobfiles/celeba_single/iti_gen/train/Young.sh. To run all of them, do the following:

find jobfiles/celeba_single/iti_gen/train/*.sh | xargs -n1 bash

Image generation

Generate images with vanilla Stable Diffusion

python models/sd/scripts/txt2img.py \
    --config='models/sd/configs/stable-diffusion/v1-inference.yaml' \
    --ckpt='models/sd/models/ldm/stable-diffusion-v1/model.ckpt' \
    --plms \
    --prompt="a headshot of a person" \
    --negative_prompt="" \
    --outdir='results/celeba_single/vanilla' \
    --skip_grid \
    --n_iter=630 \>
    --n_samples=8 \
    --seed=42

Generate images with ITI-GEN

You need to run jobfiles/celeba_single/iti_gen/generation/${attribute}.sh, for example jobfiles/celeba_single/iti_gen/generation/Young.sh. To run all of them, do the following

find jobfiles/celeba_single/iti_gen/generation/*.sh | xargs -n1 bash

Generate images with HPS

You need to run jobfiles/celeba_single/hps/generation/${attribute}.sh, for example jobfiles/celeba_single/hps/generation/Young.sh. To run all of them, do the following

find jobfiles/celeba_single/hps/generation/*.sh | xargs -n1 bash

Generate images with HPS with negative prompting

You need to run jobfiles/celeba_single/hps_negative/generation/${attribute}.sh, for example jobfiles/celeba_single/hps_negative/generation/Young.sh. To run all of them, do the following

find jobfiles/celeba_single/hps_negative/generation/*.sh | xargs -n1 bash

Evaluation

Compute FID score

Using the default parameters, 104 images are generated per category, making a total of $104 \times 80 = 8320$ images. In the original paper only 5K images are used, so we decide to take just 63 images per category, which adds up to $63 \times 80 = 5040$ images.

python reduce_num_images.py
python fid.py --ffhq --folder1 "results/celeba_single/5K_iti_gen"
python fid.py --ffhq --folder1 "results/celeba_single/5K_hps"
python fid.py --ffhq --folder1 "results/celeba_single/5K_hps_negative"
python fid.py --ffhq --folder1 "results/celeba_single/vanilla"

Compute KL-divergences

You need to run jobfiles/celeba_single/${method}/evaluation/${attribute}.sh, for example jobfiles/celeba_single/hps_negative/evaluation/Young.sh. To run all of them, do the following

find jobfiles/celeba_single -name "*.sh" | grep evaluation | xargs -n1 bash