Awesome

Enhancing Diffusion Models with Text-Encoder Reinforcement Learning

Official PyTorch codes for paper Enhancing Diffusion Models with Text-Encoder Reinforcement Learning

<a href="https://colab.research.google.com/drive/1pC6lzFR4hohwWyBgnkeZUghi2jqjl9gI?usp=sharing"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="google colab logo"></a>

Requirements & Installation

• Clone the repo and install required packages with

# git clone this repository
git clone https://github.com/chaofengc/TexForce.git
cd TexForce 

# create new anaconda env
conda create -n texforce python=3.8
source activate texforce 

# install python dependencies
pip3 install -r requirements.txt

Results on SDXL-Turbo

We also applied our method to the recent model sdxl-turbo. The model is trained with ImageReward feedback through direct back-propagation to save training time. Test with the following codes

## Note: sdturboxl requires latest diffusers installed from source with the following command
git clone https://github.com/huggingface/diffusers
cd diffusers
pip install -e .

from diffusers import AutoPipelineForText2Image
import torch

pipe = AutoPipelineForText2Image.from_pretrained("stabilityai/sdxl-turbo", torch_dtype=torch.float16, variant="fp16")
pipe = pipe.to("cuda")
pipe.load_lora_weights('chaofengc/sdxl-turbo_texforce')

pt = ['a photo of a cat.']
img = pipe(prompt=pt, num_inference_steps=1, guidance_scale=0.0).images[0]

Here are some example results:

Results on SD-Turbo

We applied our method to the recent model sdturbo. The model is trained with Q-Instruct feedback through direct back-propagation to save training time. Test with the following codes

## Note: sdturbo requires latest diffusers>=0.24.0 with AutoPipelineForText2Image class

from diffusers import AutoPipelineForText2Image
from peft import PeftModel
import torch

pipe = AutoPipelineForText2Image.from_pretrained("stabilityai/sd-turbo", torch_dtype=torch.float16, variant="fp16")
pipe = pipe.to("cuda")
PeftModel.from_pretrained(pipe.text_encoder, 'chaofengc/sd-turbo_texforce')

pt = ['a photo of a cat.']
img = pipe(prompt=pt, num_inference_steps=1, guidance_scale=0.0).images[0]

Here are some example results:

Results on SD-1.4, SD-1.5, SD-2.1

Due to code compatibility, you need to install the following diffusers first:

pip uninstall diffusers
pip install diffusers==0.16.0

You may simply load the pretrained lora weights with the following code block to improve performance of original stable diffusion model:

from diffusers import StableDiffusionPipeline
from diffusers import DDIMScheduler 
from peft import PeftModel
import torch

def load_model_weights(pipe, weight_path, model_type):
    if model_type == 'text+lora':
        text_encoder = pipe.text_encoder
        PeftModel.from_pretrained(text_encoder, weight_path)
    elif model_type == 'unet+lora':
        pipe.unet.load_attn_procs(weight_path)

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

model_id = "CompVis/stable-diffusion-v1-4"
pipe = StableDiffusionPipeline.from_pretrained(model_id, dtype=torch.float16)
pipe = pipe.to("cuda")
pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)

load_model_weights(pipe, './lora_weights/sd14_refl/', 'unet+lora')
load_model_weights(pipe, './lora_weights/sd14_texforce/', 'text+lora')

prompt = ['a painting of a dog.']
img = pipe(prompt).images[0]

Here are some example results:

Training

We rewrite the training codes based on trl with the latest diffusers library.

[!NOTE]
The latest diffusers support simple loading of lora weights with pipeline.load_lora_weights after training.

You may train the model with the following command:

Example script for single prompt training

accelerate launch --num_processes 2 src/train_ddpo.py \
    --mixed_precision="fp16" \
    --sample_num_steps 50 --train_timestep_fraction 0.5 \
    --num_epochs 40 \
    --sample_batch_size 4 --sample_num_batches_per_epoch 64 \
    --train_batch_size 4 --train_gradient_accumulation_steps 1 \
    --prompt="single" --single_prompt_type="hand" --reward_list="handdetreward" \
    --per_prompt_stat_tracking=True \
    --tracker_project_name="texforce_hand" \
    --log_with="tensorboard"

The supported prompts and reward functions are listed below:

• prompts: hand, face, color, count, comp, location
• rewards: handdetreward, topiq_nr-face, imagereward

Example script for complex multi-prompt training

accelerate launch --num_processes 2 src/train_ddpo.py \
    --mixed_precision="fp16" \
    --sample_num_steps 50 --train_timestep_fraction 0.5 \
    --num_epochs 50 \
    --sample_batch_size 4 --sample_num_batches_per_epoch 128 \
    --train_batch_size 4 --train_gradient_accumulation_steps 4 \
    --prompt="imagereward" --reward_list="imagereward" \
    --per_prompt_stat_tracking=True \
    --tracker_project_name="texforce_imgreward" \
    --log_with="tensorboard"

The supported prompts and reward functions are:

• prompts: imagereward, hps
• rewards: imagereward, hpsreward, laion_aes

Citation

If you find this code useful for your research, please cite our paper:

@inproceedings{chen2024texforce,
  title={Enhancing Diffusion Models with Text-Encoder Reinforcement Learning},
  author={Chaofeng Chen and Annan Wang and Haoning Wu and Liang Liao and Wenxiu Sun and Qiong Yan and Weisi Lin},
  year={2024},
  booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
}

License

This work is licensed under NTU S-Lab License 1.0 and a <a rel="license" href="http://creativecommons.org/licenses/by-nc-sa/4.0/">Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License</a>.

<a rel="license" href="http://creativecommons.org/licenses/by-nc-sa/4.0/"><img alt="Creative Commons License" style="border-width:0" src="https://i.creativecommons.org/l/by-nc-sa/4.0/88x31.png" /></a>

Acknowledgement

This project is largely based on trl. The hand detection codes are taken from Unified-Gesture-and-Fingertip-Detection. Many thanks to their great work :hugs:!