Add-it: Training-Free Object Insertion in Images With Pretrained Diffusion Models

Adding Object into images based on text instructions is a challenging task in semantic image editing, requiring a balance between preserving the original scene and seamlessly integrating the new object in a fitting location. Despite extensive efforts, existing models often struggle with this balance, particularly with finding a natural location for adding an object in complex scenes. We introduce Add-it, a training-free approach that extends diffusion models' attention mechanisms to incorporate information from three key sources: the scene image, the text prompt, and the generated image itself. Our weighted extended-attention mechanism maintains structural consistency and fine details while ensuring natural object placement. Without task-specific fine-tuning, Add-it achieves state-of-the-art results on both real and generated image insertion benchmarks, including our newly constructed "Additing Affordance Benchmark" for evaluating object placement plausibility, outperforming supervised methods. Human evaluations show that Add-it is preferred in over 80% of cases, and it also demonstrates improvements in various automated metrics.

EffoVPR: Effective Foundation Model Utilization for Visual Place Recognition

The task of Visual Place Recognition (VPR) is to predict the location of a query image from a database of geo-tagged images. Recent studies in VPR have highlighted the significant advantage of employing pre-trained foundation models like DINOv2 for the VPR task. However, these models are often deemed inadequate for VPR without further fine-tuning on task-specific data. In this paper, we propose a simple yet powerful approach to better exploit the potential of a foundation model for VPR. We first demonstrate that features extracted from self-attention layers can serve as a powerful re-ranker for VPR. Utilizing these features in a zero-shot manner, our method surpasses previous zero-shot methods and achieves competitive results compared to supervised methods across multiple datasets. Subsequently, we demonstrate that a single-stage method leveraging internal ViT layers for pooling can generate global features that achieve state-of-the-art results, even when reduced to a dimensionality as low as 128D. Nevertheless, incorporating our local foundation features for re-ranking, expands this gap. Our approach further demonstrates remarkable robustness and generalization, achieving state-of-the-art results, with a significant gap, in challenging scenarios, involving occlusion, day-night variations, and seasonal changes.

Unveiling the Power of Diffusion Features For Personalized Segmentation and Retrieval

Personalized retrieval and segmentation aim to locate specific instances within a dataset based on an input image and a short description of the reference instance. While supervised methods are effective, they require extensive labeled data for training. Recently, self-supervised foundation models have been introduced to these tasks showing comparable results to supervised methods. However, a significant flaw in these models is evident: they struggle to locate a desired instance when other instances within the same class are presented. In this paper, we explore text-to-image diffusion models for these tasks. Specifically, we propose a novel approach called PDM for Personalized Features Diffusion Matching, that leverages intermediate features of pre-trained text-to-image models for personalization tasks without any additional training. PDM demonstrates superior performance on popular retrieval and segmentation benchmarks, outperforming even supervised methods. We also highlight notable shortcomings in current instance and segmentation datasets and propose new benchmarks for these tasks.

Fixed-point Inversion for Text-to-image diffusion models

Text-guided diffusion models offer powerful new ways to generate and manipulate images. Several applications of these models, including image editing interpolation, and semantic augmentation, require diffusion inversion. This is the process of finding a noise seed that can be used to generate a given image. Current techniques for inverting a given image can be slow or inaccurate. The technical challenge for inverting the diffusion process arises from an implicit equation over the latent that cannot be solved in closed form. Previous approaches proposed to solve this issue by approximation or various learning schemes. Here, we formulate the problem as a fixed-point equation problem and solve it using fixed-point iterations, a well-studied approach in numerical analysis. We further identify a source of inconsistency that significantly hurts the inversion of real images encoded to the latent space. We show how to correct it by applying a prompt-aware adjustment of the encoding. Our solution, Fixed-point inversion, is much faster than previous techniques like EDICT and Null-text, with similar inversion quality. It can be combined with any pretrained diffusion model and requires no model training, prompt tuning, or additional parameters. In a series of experiments, we find that Fixed-point inversion shows improved results in several downstream tasks: image editing, image interpolation, and generation of rare objects.

Norm-guided latent space exploration for text-to-image generation

Text-to-image diffusion models show great potential in synthesizing a large variety of concepts in new compositions and scenarios. However, their latent seed space is still not well understood and has been shown to have an impact in generating new and rare concepts. Specifically, simple operations like interpolation and centroid finding work poorly with the standard Euclidean and spherical metrics in the latent space. This paper makes the observation that current training procedures make diffusion models biased toward inputs with a narrow range of norm values. This has strong implications for methods that rely on seed manipulation for image generation that can be further applied to few-shot and long-tail learning tasks. To address this issue, we propose a novel method for interpolating between two seeds and demonstrate that it defines a new non-Euclidean metric that takes into account a norm-based prior on seeds. We describe a simple yet efficient algorithm for approximating this metric and use it to further define centroids in the latent seed space. We show that our new interpolation and centroid evaluation techniques significantly enhance the generation of rare concept images. This further leads to state-of-the-art performance on few-shot and long-tail benchmarks, improving prior approach in terms of generation speed, image quality, and semantic content.

It is all about where you start: Text-to-image generation with seed selection

Text-to-image diffusion models can synthesize a large variety of concepts in new compositions and scenarios. However, they still struggle with generating uncommon concepts, rare unusual combinations, or structured concepts like hand palms. Their limitation is partly due to the long-tail nature of their training data: web-crawled data sets are strongly unbalanced, causing models to under-represent concepts from the tail of the distribution. Here we characterize the effect of unbalanced training data on text-to-image models and offer a remedy. We show that rare concepts can be correctly generated by carefully selecting suitable generation seeds in the noise space, a technique that we call SeedSelect. SeedSelect is efficient and does not require retraining the diffusion model. We evaluate the benefit of SeedSelect on a series of problems. First, in few-shot semantic data augmentation, where we generate semantically correct images for few-shot and long-tail benchmarks. We show classification improvement on all classes, both from the head and tail of the training data of diffusion models. We further evaluate SeedSelect on correcting images of hands, a well-known pitfall of current diffusion models, and show that it improves hand generation substantially.

Distributional Robustness Loss for Long-tail Learning

Real-world data is often unbalanced and long-tailed, but deep models struggle to recognize rare classes in the presence of frequent classes. To address unbalanced data, most studies try balancing the data, the loss, or the classifier to reduce classification bias towards head classes. Far less attention has been given to the latent representations learned with unbalanced data. We show that the feature extractor part of deep networks suffers greatly from this bias. We propose a new loss based on robustness theory, which encourages the model to learn high-quality representations for both head and tail classes. While the general form of the robustness loss may be hard to compute, we further derive an easy-to-compute upper bound that can be minimized efficiently. This procedure reduces representation bias towards head classes in the feature space and achieves new SOTA results on CIFAR100-LT, ImageNet-LT, and iNaturalist long-tail benchmarks. We find that training with robustness increases recognition accuracy of tail classes while largely maintaining the accuracy of head classes. The new robustness loss can be combined with various classifier balancing techniques and can be applied to representations at several layers of the deep model.

Long-tail learning with attributes

Learning to classify images with unbalanced class distributions is challenged by two effects: It is hard to learn tail classes that have few samples, and it is hard to adapt a single model to both richly-sampled and poorly-sampled classes. To address few-shot learning of tail classes, it is useful to fuse additional information in the form of semantic attributes and classify based on multi-modal information. Unfortunately, as we show below, unbalanced data leads to a "familiarity bias", where classifiers favor sample-rich classes. This bias and lack of calibrated predictions make it hard to fuse correctly information from multiple modalities like vision and attributes. Here we describe DRAGON, a novel modular architecture for long-tail learning designed to address these biases and fuse multi-modal information in face of unbalanced data. Our architecture is based on three classifiers: a vision expert, a semantic attribute expert that excels on the tail classes, and a debias-and-fuse module to combine their predictions. We present the first benchmark for long-tail learning with attributes and use it to evaluate DRAGON. DRAGON outperforms state-of-the-art long-tail learning models and Generalized Few-Shot-Learning with attributes (GFSL-a) models. DRAGON also obtains SoTA in some existing benchmarks for single-modality GFSL.