CleanDIFT: Diffusion Features without Noise

Internal features from large-scale pre-trained diffusion models have recently been established as powerful semantic descriptors for a wide range of downstream tasks. Works that use these features generally need to add noise to images before passing them through the model to obtain the semantic features, as the models do not offer the most useful features when given images with little to no noise. We show that this noise has a critical impact on the usefulness of these features that cannot be remedied by ensembling with different random noises. We address this issue by introducing a lightweight, unsupervised fine-tuning method that enables diffusion backbones to provide high-quality, noise-free semantic features. We show that these features readily outperform previous diffusion features by a wide margin in a wide variety of extraction setups and downstream tasks, offering better performance than even ensemble-based methods at a fraction of the cost.

Distillation of Diffusion Features for Semantic Correspondence

Semantic correspondence, the task of determining relationships between different parts of images, underpins various applications including 3D reconstruction, image-to-image translation, object tracking, and visual place recognition. Recent studies have begun to explore representations learned in large generative image models for semantic correspondence, demonstrating promising results. Building on this progress, current state-of-the-art methods rely on combining multiple large models, resulting in high computational demands and reduced efficiency. In this work, we address this challenge by proposing a more computationally efficient approach. We propose a novel knowledge distillation technique to overcome the problem of reduced efficiency. We show how to use two large vision foundation models and distill the capabilities of these complementary models into one smaller model that maintains high accuracy at reduced computational cost. Furthermore, we demonstrate that by incorporating 3D data, we are able to further improve performance, without the need for human-annotated correspondences. Overall, our empirical results demonstrate that our distilled model with 3D data augmentation achieves performance superior to current state-of-the-art methods while significantly reducing computational load and enhancing practicality for real-world applications, such as semantic video correspondence. Our code and weights are publicly available on our project page.

Scene Graph Conditioning in Latent Diffusion

Diffusion models excel in image generation but lack detailed semantic control using text prompts. Additional techniques have been developed to address this limitation. However, conditioning diffusion models solely on text-based descriptions is challenging due to ambiguity and lack of structure. In contrast, scene graphs offer a more precise representation of image content, making them superior for fine-grained control and accurate synthesis in image generation models. The amount of image and scene-graph data is sparse, which makes fine-tuning large diffusion models challenging. We propose multiple approaches to tackle this problem using ControlNet and Gated Self-Attention. We were able to show that using out proposed methods it is possible to generate images from scene graphs with much higher quality, outperforming previous methods. Our source code is publicly available on https://github.com/FrankFundel/SGCond

Automatic Bat Call Classification using Transformer Networks

Automatically identifying bat species from their echolocation calls is a difficult but important task for monitoring bats and the ecosystem they live in. Major challenges in automatic bat call identification are high call variability, similarities between species, interfering calls and lack of annotated data. Many currently available models suffer from relatively poor performance on real-life data due to being trained on single call datasets and, moreover, are often too slow for real-time classification. Here, we propose a Transformer architecture for multi-label classification with potential applications in real-time classification scenarios. We train our model on synthetically generated multi-species recordings by merging multiple bats calls into a single recording with multiple simultaneous calls. Our approach achieves a single species accuracy of 88.92% (F1-score of 84.23%) and a multi species macro F1-score of 74.40% on our test set. In comparison to three other tools on the independent and publicly available dataset ChiroVox, our model achieves at least 25.82% better accuracy for single species classification and at least 6.9% better macro F1-score for multi species classification.