UniFL: Improve Stable Diffusion via Unified Feedback Learning

Text-to-Video Examples

AnimateDiff + LCM 12-Step

AnimateDiff + UniFL 12-Step

AnimateDiff + LCM 12-Step

AnimateDiff + UniFL 12-Step

"a drop of black ink falling in a superwhite background in slow motions angle advertising."

"3d animated female character, singing and moving her arms, musical film style , 4k, high quality vid ..."

":lipsticks advertisement -camera pan up right -s 7."

"modern swimming pool villa design by arquitetura, in winter, snow, in the style of 32k uhd, white and blue ... "

"modern swimming pool villa design by arquitetura, in winter, snow, in the style of 32k uhd, white and blue ..."

"A cute talking panda stands in the museum, furry，Chinese New Year style red clothes， cinematic, 4k, epic Steven Spielberg movie still ..."

"a uniquely captivating Ferrari Portofino advertisement, emphasizing its elegance and speed, night, ultra-high definition."

Abstract

Diffusion models have revolutionized the field of image generation, leading to the proliferation of high-quality models and diverse downstream applications. However, despite these significant advancements, the current competitive solutions still suffer from several limitations, including inferior visual quality, a lack of aesthetic appeal, and inefficient inference, without a comprehensive solution in sight. To address these challenges, we present UniFL, a unified framework that leverages feedback learning to enhance diffusion models comprehensively. UniFL stands out as a universal, effective, and generalizable solution applicable to various diffusion models, such as SD1.5 and SDXL. Notably, UniFL incorporates three key components: perceptual feedback learning, which enhances visual quality; decoupled feedback learning, which improves aesthetic appeal; and adversarial feedback learning, which optimizes inference speed. In-depth experiments and extensive user studies validate the superior performance of our proposed method in enhancing both the quality of generated models and their acceleration. For instance, UniFL surpasses ImageReward by 17% user preference in terms of generation quality and outperforms LCM and SDXL Turbo by 57% and 20% in 4-step inference. Moreover, we have verified the efficacy of our approach in downstream tasks, including Lora, ControlNet, and AnimateDiff.

Perceptual Feedback Learning

The existing perceptual model can serve as an excellent visual generation quality feedback provider for the diffusion model. Take the instance segmentation model as an example, it accurately captures the defect of the generation(e.g. the distort arm of the boy)

Perceptual Feedback Learning of UniFL significantly enhance the visual generation quality of diffusion model, including style response, structure optimization.

Adversarial Feedback Learning

Adversarial Feedback Learning of UniFL achieves superior acceleration performance compared with the existing commonly used acceleration methods, including LCM, SDXL-Turbo, etc.

Comparsions

UniFL exhibits a remarkable advantage over existing methods that concentrate on quality optimization and inference acceleration in terms of both quantitative comparison and user study.

BibTeX

@misc{zhang2024unifl,
      title={UniFL: Improve Stable Diffusion via Unified Feedback Learning}, 
      author={Jiacheng Zhang and Jie Wu and Yuxi Ren and Xin Xia and Huafeng Kuang and Pan Xie and Jiashi Li and Xuefeng Xiao and Weilin Huang and Min Zheng and Lean Fu and Guanbin Li},
      year={2024},
      eprint={2404.05595},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

UniFL:

Improve Stable Diffusion via Unified Feedback Learning

Text-to-Image Examples