tgxs002 / align_sd Goto Github PK

select_training_images.py gives 42201 non-preferred images, which is inconsistent with 21108 mentioned in paper. Is there any other parameter that can be used to exactly reproduce the count of preferred and non-preferred images?

Hey, thanks for the great work!

I am interested in reproducing the numbers from table 2 in your paper. Could you please advise on how to do that? Can I directly use the test.json provided in your repo? What exact metric do you use for these results, it wasn't directly clear from the paper. Thanks!

Hi, have you tested on stable diffusion 2.1?

Thanks for your excellent work! I wonder know how you set your accelerate config , like if you use deepspeed .since I tried to use it, but unfortunately I failed. I would be grateful if you could answer my question!

Nice work, but when will the training code be released? I'm hoping for it.

Hi, I have noticed you release adapted_model.bin in V1 align_sd. I wonder if threre is a new adapted_model in V2?

Dear Authors,

Thank you for your awesome work! Could you please add a license to your repo?

Hi, great research! Impressed by the results.

For possibly your own interest, and in case anybody else come across this, you can use this conversion script to get the LoRA models functioning with AUTOMATIC1111/stable-diffusion-webui, the interface majority of the SD community uses. Credit to harrywang for the original script.

import re
import os
import argparse
import torch
from safetensors.torch import save_file

def main(args):
    if torch.cuda.is_available():
        device = 'cuda'
        checkpoint = torch.load(args.file, map_location=torch.device('cuda'))
    else:
        device = 'cpu'
        checkpoint = torch.load(args.file, map_location=torch.device('cpu'))
    
    new_dict = dict()
    for idx, key in enumerate(checkpoint):
        new_key = re.sub('\.processor\.', '_', key)
        new_key = re.sub('mid_block\.', 'mid_block_', new_key)
        new_key = re.sub('_lora.up.', '.lora_up.', new_key)
        new_key = re.sub('_lora.down.', '.lora_down.', new_key)
        new_key = re.sub('\.(\d+)\.', '_\\1_', new_key)
        new_key = re.sub('to_out', 'to_out_0', new_key)
        new_key = 'lora_unet_' + new_key

        new_dict[new_key] = checkpoint[key]

    file_name = os.path.splitext(args.file)[0]
    new_lora_name = file_name + '_converted.safetensors'
    print("Saving " + new_lora_name)
    save_file(new_dict, new_lora_name)

def parse_args():
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--file",
        type=str,
        default=None,
        required=True,
    )
    
    args = parser.parse_args()
    return args

if __name__ == "__main__":
    args = parse_args()
    main(args)

Thank you, I have successfully replicated the training of LoRA according to your code, and it does improve the performance of Stable Diffusion significantly. May I ask when the training code for hps classifier will be released? @tgxs002

Could the authors release the validation_prompt.json file? Thus we can repeat the visualization results and make a comparison with the results reported in the paper.

Thanks

I am very interested in your work and currently attempting to reproduce your results. I ran the script download_regularization_images.py to download the provided regularization_images data, which consists of image-text pairs. I would like to know how to preprocess it to incorporate it into the LoRA training.Thanks. @tgxs002

It seems to be having trouble "building".

Thanks for your great work!
I wonder how much time it costs to train the Lora on the dataset (1M DiffusionDB + subset of Laion5B) on, for example, 4 GPUs?

The training dataset for hps classifier is too large to download. I have tried to download it many times, but all attempts have failed. Would you be able to provide an alternative download link, such as Baidu Cloud or another platform? @tgxs002

Hello, I am very interested in your work and I am trying to reproduce your results. Would it be possible for you to share positive and negative sample data you used for LoRA training? @tgxs002

Hi @tgxs002 , thanks for your work, and making the dataset and classifier open-sourced!

As a sanity check, I evaluated your trained HPC on the examples in the training data that are preferred by humans (S1), and the examples in the training data that are unpreferred by humans (S2).

I found that the average HPS in the setting S1 ~ 21.0 whereas the HPC in the setting S2 ~ 20.26. For a good classifier, I was hoping that the scores in the setting S2 be very low as compared to S1, but it is not the case. Does it mean that the HPC is not trained properly, but it seems contradictory because the paper claims that the HPC has good agreement with humans?

And if my evaluation numbers look too off, can you let me know what you are getting at your end?

Thank you for your great work!
I noticed that you trained the lora model based on 'CompVis/stable-diffusion-v1-4' and I want to know if I need to retrain the lora model if I have a different SD model?

Hi ! Thanks for sharing this dataset :)

I was wondering if you planned to share the dataset on Hugging Face as well ? This way anyone in the research community can visualise it online and even load it in one line of python

This is a nice tool. As far as I have understood, it generates sequence of images based on text. But,

for the case, where I have sequence of images - frames and some sentences. Is it possible to match the sequence based on sentence text? If there is any way.