# Attention Calibration for Disentangled Text-to-Image Personalization

Source: [Attention Calibration for Disentangled Text-to-Image Personalization](<https://www.sota2.com/research/paper/68fa72dd1e209131cea5906e>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-01-02T05:41:16.169Z

Authors: Yanbing Zhang, Mengping Yang, Qin Zhou, Zhe Wang

Published: 2024-03-27T13:31:39.000Z

## Abstract

Recent thrilling progress in large-scale text-to-image (T2I) models has unlocked unprecedented synthesis quality of AI-generated content (AIGC) including image generation, 3D and video composition. Further, personalized techniques enable appealing customized production of a novel concept given only several images as reference. However, an intriguing problem persists: Is it possible to capture multiple, novel concepts from one single reference image? In this paper, we identify that existing approaches fail to preserve visual consistency with the reference image and eliminate cross-influence from concepts. To alleviate this, we propose an attention calibration mechanism to improve the concept-level understanding of the T2I model. Specifically, we first introduce new learnable modifiers bound with classes to capture attributes of multiple concepts. Then, the classes are separated and strengthened following the activation of the cross-attention operation, ensuring comprehensive and self-contained concepts. Additionally, we suppress the attention activation of different classes to mitigate mutual influence among concepts. Together, our proposed method, dubbed DisenDiff, can learn disentangled multiple concepts from one single image and produce novel customized images with learned concepts. We demonstrate that our method outperforms the current state of the art in both qualitative and quantitative evaluations. More importantly, our proposed techniques are compatible with LoRA and inpainting pipelines, enabling more interactive experiences.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Multi-Concept Personalization on Cat+Dog](<https://www.sota2.com/research/sota/multi-concept-personalization-on-cat-plus-dog.md>): Multi-Concept Personalization; Cat+Dog; Image Alignment (Mean): 82.4.

- [Multi-Concept Personalization on Cow+Bird](<https://www.sota2.com/research/sota/multi-concept-personalization-on-cow-plus-bird.md>): Multi-Concept Personalization; Cow+Bird; Image Alignment (Mean): 78.3.

- [Multi-Concept Personalization on Man+Woman](<https://www.sota2.com/research/sota/multi-concept-personalization-on-man-plus-woman.md>): Multi-Concept Personalization; Man+Woman; Image-alignment (Mean): 0.749.

- [Multi-Concept Personalization on Chair+Vase](<https://www.sota2.com/research/sota/multi-concept-personalization-on-chair-plus-vase.md>): Multi-Concept Personalization; Chair+Vase; Image Alignment (Mean): 82.2.

- [Multi-Concept Personalization on Dog+Pig](<https://www.sota2.com/research/sota/multi-concept-personalization-on-dog-plus-pig.md>): Multi-Concept Personalization; Dog+Pig; Image-alignment (Mean): 77.3.

- [Multi-Concept Personalization on Mother+Child](<https://www.sota2.com/research/sota/multi-concept-personalization-on-mother-plus-child.md>): Multi-Concept Personalization; Mother+Child; Image Alignment (Mean): 0.718.

- [Multi-Concept Personalization on Woman+Dog](<https://www.sota2.com/research/sota/multi-concept-personalization-on-woman-plus-dog.md>): Multi-Concept Personalization; Woman+Dog; Image Alignment (Mean): 73.7.

- [Multi-Concept Personalization on Baby+Toy](<https://www.sota2.com/research/sota/multi-concept-personalization-on-baby-plus-toy.md>): Multi-Concept Personalization; Baby+Toy; Image Alignment (Mean): 0.808.

- [Multi-Concept Personalization on Chair+Lamp](<https://www.sota2.com/research/sota/multi-concept-personalization-on-chair-plus-lamp.md>): Multi-Concept Personalization; Chair+Lamp; Image Alignment (Mean): 79.5.

- [Multi-Concept Personalization on Horse+Dog](<https://www.sota2.com/research/sota/multi-concept-personalization-on-horse-plus-dog.md>): Multi-Concept Personalization; Horse+Dog; Image Alignment (Mean): 74.4.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2403.18551>)

[PDF](<https://arxiv.org/pdf/2403.18551.pdf>)
