kousw / experimental-dreamtuner

an unofficial implementation of dreamtuner

About

I have conducted an experimental implementation of DreamTuner (https://arxiv.org/abs/2312.13691). This is an unofficial implementation, and while it is unclear if the implementation is correct, some positive effects have been observed.

The implementation involves a two-step training process:

1. Subject Encoder Training
2. Subject Driven Fine-tuning

In the Subject Encoder Training, a Subject Encoder is used for feature extraction of reference images. A Self-Encoder-Attention layer is added between the Self-Attention and Cross-Attention layers of the Unet's Transformer Block, and training is conducted (see train.py). Inference using only the Subject Encoder can be performed using inference.py.

Subject Driven Fine-tuning essentially involves fine-tuning through Dreambooth + Textual Inversion. Prior to this implementation, the effects of Self-Subject Attention and the implementation of a Unet and Reference Guidance (as a replacement for CFG) for reference images need to be evaluated. Self-Subject Driven inference can be performed using inference_ss.py.

If the generated results are satisfactory, Subject Driven Fine-tuning is carried out (see train_dreambooth.py). The resulting Subject Encoder, Unet, and Text Inversion models can be used for inference with inference_ss.py.

Currently, there is no support for Subject Driven Inference using ControlNet.

Note: Some aspects of the implementation diverge from the details in the paper. Particularly, the Subject Encoder showed no significant difference, so it has been replaced with the SSR-Encoder implementation (https://arxiv.org/abs/2312.16272).

Training was conducted on an anime model with about 50,000 images. As described in the paper, some effects were anticipated, but due to the quality and quantity of the dataset and GPU limitations leading to a higher learning rate, the results were not as responsive to prompts as expected. If you are experimenting, it's necessary to properly adjust values such as the Subject Encoder beta and the Reference Guidance Scale.

Considering the GPU processing time and the dataset's quality and quantity, further training by an individual is deemed challenging. While I do not intend to continue development, the implementation is made public.

The trained Subject Encoder can be downloaded from here . Replace the Pretrained models under models/dreamtuner with it for use.

Environment

Python 3.10.9 CUDA 12.2

Installation

python -m venv venv
source venv/bin/activate

pip install -r requirements.txt

Downloading the Base Model

mkdir datasets
mkdir models

cd ./models
git clone some_diffusers_sd_15_model_from_huggin_face
# ex) git clone https://huggingface.co/runwayml/stable-diffusion-v1-5 or some_anime_model

# copy base model for finetuning
cp -r some_diffusers_sd_15_model_from_huggin_face/ dreamtuner

Preparing the Dataset

# your datasets into ./datasets 

Creating images for background removal, masks, and depth

python process.py path

Creating tags

python process_tags.py path

Training

with accelerater

# configulation
accelerate config

# training 
bash ./train.sh

When training is complete, copy the generated files (subject_encoder, unet) to 'models/dreamtuner' folder.

You will need to modify model_index.json as follows to replace the subject encoder and unet modules when loading. Also, please modify it if you have a different scheduler, as it will not work with anything other than the DDIM scheduler.

{
    "_class_name": "DreamTunerPipelineSelfSubject",
    "_diffusers_version": "0.8.0.dev0",
    "feature_extractor": [
      "transformers",
      "CLIPImageProcessor"
    ],
    "safety_checker": [
      "stable_diffusion",
      "StableDiffusionSafetyChecker"
    ],
    "scheduler": [
      "diffusers",
      "DDIMScheduler"
    ],
    "text_encoder": [
      "transformers",
      "CLIPTextModel"
    ],
    "tokenizer": [
      "transformers",
      "CLIPTokenizer"
    ],
    "subject_encoder": [
      "dreamtuner.models.subject_encoder",
      "SubjectEncoder"
    ],
    "unet": [
      "dreamtuner.models.unet",
      "SDUNet2DConditionModel"
    ],
    "vae": [
      "diffusers",
      "AutoencoderKL"
    ]
}

train dreambooth

bash ./train_dreambooth.sh

Inference

Subject Encoder inference

python inference.py --model_name_or_path ./models/dreamtuner --subject_encoder_beta 0.5 --num_samples 1 --num_inference_steps 50 --reference_image ./datasets/sample/00006_rgb.png --negative_prompt "worst quality, normal quality, low quality, low res, blurry, text, watermark, logo, banner, extra digits, cropped, jpeg artifacts, signature, username, error, sketch ,duplicate, ugly, monochrome, horror, geometry, mutation, disgusting" --prompt "best quality,1girl,outdoor"

Subject Driven inference

witout mask, without reference guidance (self-subject attention only)

python inference_ss.py --model_name_or_path ./models/dreamtuner --subject_encoder_beta 0.2 --num_samples 1 --num_inference_steps 50 --reference_image ./datasets/sample/00006_rgb.png --negative_prompt "worst quality, normal quality, low quality, low res, blurry, text, watermark, logo, banner, extra digits, cropped, jpeg artifacts, signature, username, error, sketch ,duplicate, ugly, monochrome, horror, geometry, mutation, disgusting" --prompt "best quality,1girl,outdoor" --dtype float16

with mask, with reference guidance

python inference_ss.py --model_name_or_path ./models/dreamtuner --subject_encoder_beta 0.2 --num_samples 1 --num_inference_steps 50 --reference_image ./datasets/sample/00006_rgb.png --mask_image ./datasets/sample/00006_mask.png --negative_prompt "worst quality, normal quality, low quality, low res, blurry, text, watermark, logo, banner, extra digits, cropped, jpeg artifacts, signature, username, error, sketch ,duplicate, ugly, monochrome, horror, geometry, mutation, disgusting" --prompt "best quality,1girl,outdoor" --enable_reference_guidance --reference_guidance_scale 2  --dtype float16

with textual invertion

python inference_ss.py --model_name_or_path ./models/dreamtuner_tuned --text_embeds_name_or_path ./models/dreamtuner_tuned/textual_inversion --subject_encoder_beta 0.2 --num_samples 1 --num_inference_steps 50 --reference_image ./datasets/sample/00009_rgb.png --mask_image ./datasets/sample/00009_mask.png --negative_prompt "worst quality, normal quality, low quality, low res, blurry, text, watermark, logo, banner, extra digits, cropped, jpeg artifacts, signature, username, error, sketch ,duplicate, ugly, monochrome, horror, geometry, mutation, disgusting" --prompt "best quality,1girl,outdoor" --enable_reference_guidance --reference_guidance_scale 2  --dtype float16

Sample

Original

Generated Images

experimental-dreamtuner's Issues

I get some bad result by following your step

I download you provioded checkpoint, which contain Unet and subject-encoder.
Then, I excute this comand line:

python inference_ss.py --model_name_or_path ./models/dreamtuner --subject_encoder_beta 0.2 --num_samples 10 --num_inference_steps 50
--reference_image ./datasets/sample/00006_rgb.png
--mask_image ./datasets/sample/00006_mask.png
--negative_prompt "worst quality, normal quality, low quality, low res, blurry, text, watermark, logo, banner, extra digits, cropped, jpeg artifacts, signature, username, error, sketch ,duplicate, ugly, monochrome, horror, geometry, mutation, disgusting"
--prompt "best quality,1girl,outdoor" --enable_reference_guidance --reference_guidance_scale 2 --dtype float16

But I get these result:

I use "datasets/sample/00006_rgb.png" as reference image.
The output image quality is worse than one generated by using "inference.py"

by the way, all of result is not like the reference image

At last, can you provide the script which make DreamBooth dataset.

can it generate different pose for one anime person?

Can you give me more generative examples?
from given generative example, the role's face pose is fixed. I am confused whether pose can be changed?
look forward to receive your message.