mapo-t2i/mapo-beta

Margin-aware Preference Optimization for Aligning Diffusion Models without Reference

<div align="center"> <img src="https://github.com/mapo-t2i/mapo/blob/main/assets/mapo_overview.png?raw=true" width=750/> </div><br>

We propose MaPO, a reference-free, sample-efficient, memory-friendly alignment technique for text-to-image diffusion models. For more details on the technique, please refer to our paper here.

Developed by

• —Jiwoo Hong^* (KAIST AI)
• —Sayak Paul^* (Hugging Face)
• —Noah Lee (KAIST AI)
• —Kashif Rasul (Hugging Face)
• —James Thorne (KAIST AI)
• —Jongheon Jeong (Korea University)

Dataset

This model was fine-tuned from Stable Diffusion XL on the yuvalkirstain/pickapic_v2 dataset.

Training Code

Refer to our code repository here.

Qualitative Comparison

<div align="center"> <img src="assets/comparison.png" width=750/> </div>

Results

Below we report some quantitative metrics and use them to compare MaPO to existing models:

<style> table { width: 100%; border-collapse: collapse; } th, td { border: 1px solid #000; padding: 8px; text-align: center; } th { background-color: #808080; } .ours { font-style: italic; } </style>

<table> <caption>Average score for Aesthetic, HPS v2.1, and PickScore</caption> <thead> <tr> <th></th> <th>Aesthetic</th> <th>HPS v2.1</th> <th>Pickscore</th> </tr> </thead> <tbody> <tr> <td>SDXL</td> <td>6.03</td> <td>30.0</td> <td>22.4</td> </tr> <tr> <td>SFT_Chosen</td> <td>5.95</td> <td>29.6</td> <td>22.0</td> </tr> <tr> <td>Diffusion-DPO</td> <td>6.03</td> <td>31.1</td> <td><b>22.7<b/></td> </tr> <tr> <td><b>MaPO (Ours)<b/></td> <td><b>6.17<b/></td> <td><b>31.2<b/></td> <td>22.5</td> </tr> </tbody> </table>

We evaluated this checkpoint in the Imgsys public benchmark. MaPO was able to outperform or match 21 out of 25 state-of-the-art text-to-image diffusion models by ranking 7th on the leaderboard at the time of writing, compared to Diffusion-DPO’s 20th place, while also consuming 14.5% less wall-clock training time on adapting Pick-a-Pic v2. We appreciate the imgsys team for helping us get the human preference data.

<div align="center"> <img src="https://mapo-t2i.github.io/static/images/imgsys.png" width=750/> </div>

The table below reports memory efficiency of MaPO, making it a better alternative for alignment fine-tuning of diffusion models:

<table> <caption>Computational costs of Diffusion-DPO and MaPO</caption> <thead> <tr> <th></th> <th>Diffusion-DPO</th> <th>MaPO <span class="ours">(Ours)</span></th> </tr> </thead> <tbody> <tr> <td>Time (↓)</td> <td>63.5</td> <td><b>54.3 (-14.5%)</b></td> </tr> <tr> <td>GPU Mem. (↓)</td> <td>55.9</td> <td><b>46.1 (-17.5%)</b></td> </tr> <tr> <td>Max Batch (↑)</td> <td>4</td> <td><b>16 (×4)</b></td> </tr> </tbody> </table>

Inference

from diffusers import DiffusionPipeline, AutoencoderKL, UNet2DConditionModel
import torch 

sdxl_id = "stabilityai/stable-diffusion-xl-base-1.0"
vae_id = "madebyollin/sdxl-vae-fp16-fix"
unet_id = "mapo-t2i/mapo-beta"

vae = AutoencoderKL.from_pretrained(vae_id, torch_dtype=torch.float16)
unet = UNet2DConditionModel.from_pretrained(unet_id, torch_dtype=torch.float16)
pipeline = DiffusionPipeline.from_pretrained(sdxl_id, vae=vae, unet=unet, torch_dtype=torch.float16).to("cuda")

prompt = "An abstract portrait consisting of bold, flowing brushstrokes against a neutral background."
image = pipeline(prompt=prompt, num_inference_steps=30).images[0]

For qualitative results, please visit our project website.

Citation

@misc{hong2024marginaware,
  title={Margin-aware Preference Optimization for Aligning Diffusion Models without Reference}, 
  author={Jiwoo Hong and Sayak Paul and Noah Lee and Kashif Rasul and James Thorne and Jongheon Jeong},
  year={2024},
  eprint={2406.06424},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}