Per-Query Visual Concept Learning

Per-Query Visual Concept Learning

1Bar-Ilan University, 2OriginAI, 3NVIDIA

Given a pretrained text-to-image personalization checkpoint, our method enhances per-query generation in terms of image alignment and text alignment, with just a single gradient update (~4 seconds on an NVIDIA H100 GPU). It is compatible with a wide range of personalization techniques (e.g., DreamBooth, LoRA, Textual Inversion, DBlend) and supports various diffusion backbones, including UNet-based models (e.g., SDXL, SD) and transformer-based models (e.g., FLUX, SD3).

Abstract

Visual concept learning, also known as Text-to-image personalization, is the process of teaching new concepts to a pretrained model. This has numerous applications from product placement to entertainment and personalized design. Here we show that many existing methods can be substantially augmented by adding a personalization step that is (1) specific to the prompt and noise seed, and (2) using two loss terms based on the self- and cross- attention, capturing the identity of the personalized concept. Specifically, we leverage PDM features - previously designed to capture identity - and show how they can be used to improve personalized semantic similarity. We evaluate the benefit that our method gains on top of six different personalization methods, and several base text-to-image models (both UNet- and DiT-based). We find significant improvements even over previous per-query personalization methods.

Qualitative Comparisons

How does our method work?

We enhance personalized text-to-image models by computing self- and cross-attention features from a single denoising step of a generated image \( x^{\text{gen}} \) and a reference image \( x^{\text{ref}} \). Using DIFT, we match these features and define losses \( \mathcal{L}_{\text{SA}} \), \( \mathcal{L}_{\text{CA}} \), and \( \mathcal{L}_{\text{LDM}} \), which are combined to update the personalization tuning parameters via gradient descent.

Quantitative Comparisons

Effect of adding our method to baseline personalization models. shown are image alignment (left) and text alignment (right) of several personalization approaches with (w/) and without (w/o) the integration of our method.

Quantitative Comparison with previous per-query methods. Image alignment (left) and text alignment (right) of various personalization approaches (including, DB, LoRA, and TI), with (w/) and without (w/o) the integration of different per-query methods (including, AlignIT, PALP, and Ours).

Ablation Studies and Sensitivity to Parameters

Quality-vs-time Tradeoff.    Figures show the CLIP-I (left) and CLIP-T (right) metrics as a function of fine-tuning duration. The duration is longer when using more features that are collected throughout the denoising path. \( T \) is the number of feature maps using to compute the losses.

Quality-vs-time Tradeoff.    Figures show the CLIP-I (left) and CLIP-T (right) metrics as a function of fine-tuning duration. The duration is longer when using more SA features that are collected throughout the denoising path. \( T \) is the number of feature maps using to compute the losses.

Noise weight Ablation Study.    Sensitivity of CLIP-I and CLIP-T metrics to the magnitude of the noise, in terms of the \( t \) parameter used when calculating features.

Per-loss contribution ablation study.    We investigate the contribution of each loss in our proposed method. Testing across different methods, leveraging SD1.5 as backbone, and using a fixed \( \lambda_{\text{PDM}}=1,\lambda_{\text{CA}}=1 \) values. \( \mathcal{L}_{\text{PDM}} \) provides superior subject alignment, and \( \mathcal{L}_{\text{CA}} \) results superior prompt-adherence. Incorporating both provides superior results in both aspects.

Stage-wise loss effectiveness ablation study.    We examine the effectiveness of \( \mathcal{L}_{\text{SA}} \) and \( \mathcal{L}_{\text{CA}} \) throughout training rather than post-training, with two prominent methods, utilizing SD1.5 as backbone, with \( \lambda_{\text{PDM}}=1,\lambda_{\text{CA}}=1 \). We observe inferior results compared to post-training integration, and equal/inferior results compared to the baseline. This quantitatively support that certain losses become effective only after the model reaches a certain state.


BibTeX

If you find our work useful, please cite our paper:

    
    @article{malca2024perquery,
      title   = {Per-Query Visual Concept Learning},
      author  = {Malca, Ori and Samuel, Dvir and Chechik, Gal},
      journal = {arXiv preprint arXiv:2508.09045},
      year    = {2026}
    }