Concept Steerers: Leveraging K-Sparse Autoencoders for Controllable Generations

Despite the remarkable progress in text-to-image generative models, they are prone to adversarial attacks and inadvertently generate unsafe, unethical content. Existing approaches often rely on fine-tuning models to remove specific concepts, which is computationally expensive, lack scalability, and/or compromise generation quality. In this work, we propose a novel framework leveraging k-sparse autoencoders (k-SAEs) to enable efficient and interpretable concept manipulation in diffusion models. Specifically, we first identify interpretable monosemantic concepts in the latent space of text embeddings and leverage them to precisely steer the generation away or towards a given concept (e.g., nudity) or to introduce a new concept (e.g., photographic style). Through extensive experiments, we demonstrate that our approach is very simple, requires no retraining of the base model nor LoRA adapters, does not compromise the generation quality, and is robust to adversarial prompt manipulations. Our method yields an improvement of $\mathbf{20.01\%}$ in unsafe concept removal, is effective in style manipulation, and is $\mathbf{\sim5}$x faster than current state-of-the-art.

DebiasPI: Inference-time Debiasing by Prompt Iteration of a Text-to-Image Generative Model

Ethical intervention prompting has emerged as a tool to counter demographic biases of text-to-image generative AI models. Existing solutions either require to retrain the model or struggle to generate images that reflect desired distributions on gender and race. We propose an inference-time process called DebiasPI for Debiasing-by-Prompt-Iteration that provides prompt intervention by enabling the user to control the distributions of individuals' demographic attributes in image generation. DebiasPI keeps track of which attributes have been generated either by probing the internal state of the model or by using external attribute classifiers. Its control loop guides the text-to-image model to select not yet sufficiently represented attributes, With DebiasPI, we were able to create images with equal representations of race and gender that visualize challenging concepts of news headlines. We also experimented with the attributes age, body type, profession, and skin tone, and measured how attributes change when our intervention prompt targets the distribution of an unrelated attribute type. We found, for example, if the text-to-image model is asked to balance racial representation, gender representation improves but the skin tone becomes less diverse. Attempts to cover a wide range of skin colors with various intervention prompts showed that the model struggles to generate the palest skin tones. We conducted various ablation studies, in which we removed DebiasPI's attribute control, that reveal the model's propensity to generate young, male characters. It sometimes visualized career success by generating two-panel images with a pre-success dark-skinned person becoming light-skinned with success, or switching gender from pre-success female to post-success male, thus further motivating ethical intervention prompting with DebiasPI.

$\textit{Revelio}$: Interpreting and leveraging semantic information in diffusion models

We study $\textit{how}$ rich visual semantic information is represented within various layers and denoising timesteps of different diffusion architectures. We uncover monosemantic interpretable features by leveraging k-sparse autoencoders (k-SAE). We substantiate our mechanistic interpretations via transfer learning using light-weight classifiers on off-the-shelf diffusion models' features. On $4$ datasets, we demonstrate the effectiveness of diffusion features for representation learning. We provide in-depth analysis of how different diffusion architectures, pre-training datasets, and language model conditioning impacts visual representation granularity, inductive biases, and transfer learning capabilities. Our work is a critical step towards deepening interpretability of black-box diffusion models. Code and visualizations available at: https://github.com/revelio-diffusion/revelio

Mitigating stereotypical biases in text to image generative systems

State-of-the-art generative text-to-image models are known to exhibit social biases and over-represent certain groups like people of perceived lighter skin tones and men in their outcomes. In this work, we propose a method to mitigate such biases and ensure that the outcomes are fair across different groups of people. We do this by finetuning text-to-image models on synthetic data that varies in perceived skin tones and genders constructed from diverse text prompts. These text prompts are constructed from multiplicative combinations of ethnicities, genders, professions, age groups, and so on, resulting in diverse synthetic data. Our diversity finetuned (DFT) model improves the group fairness metric by 150% for perceived skin tone and 97.7% for perceived gender. Compared to baselines, DFT models generate more people with perceived darker skin tone and more women. To foster open research, we will release all text prompts and code to generate training images.

Helping Visually Impaired People Take Better Quality Pictures

Perception-based image analysis technologies can be used to help visually impaired people take better quality pictures by providing automated guidance, thereby empowering them to interact more confidently on social media. The photographs taken by visually impaired users often suffer from one or both of two kinds of quality issues: technical quality (distortions), and semantic quality, such as framing and aesthetic composition. Here we develop tools to help them minimize occurrences of common technical distortions, such as blur, poor exposure, and noise. We do not address the complementary problems of semantic quality, leaving that aspect for future work. The problem of assessing and providing actionable feedback on the technical quality of pictures captured by visually impaired users is hard enough, owing to the severe, commingled distortions that often occur. To advance progress on the problem of analyzing and measuring the technical quality of visually impaired user-generated content (VI-UGC), we built a very large and unique subjective image quality and distortion dataset. This new perceptual resource, which we call the LIVE-Meta VI-UGC Database, contains $40$K real-world distorted VI-UGC images and $40$K patches, on which we recorded $2.7$M human perceptual quality judgments and $2.7$M distortion labels. Using this psychometric resource we also created an automatic blind picture quality and distortion predictor that learns local-to-global spatial quality relationships, achieving state-of-the-art prediction performance on VI-UGC pictures, significantly outperforming existing picture quality models on this unique class of distorted picture data. We also created a prototype feedback system that helps to guide users to mitigate quality issues and take better quality pictures, by creating a multi-task learning framework.

Beyond web-scraping: Crowd-sourcing a geographically diverse image dataset

Current dataset collection methods typically scrape large amounts of data from the web. While this technique is extremely scalable, data collected in this way tends to reinforce stereotypical biases, can contain personally identifiable information, and typically originates from Europe and North America. In this work, we rethink the dataset collection paradigm and introduce GeoDE, a geographically diverse dataset with 61,940 images from 40 classes and 6 world regions, and no personally identifiable information, collected through crowd-sourcing. We analyse GeoDE to understand differences in images collected in this manner compared to web-scraping. Despite the smaller size of this dataset, we demonstrate its use as both an evaluation and training dataset, highlight shortcomings in current models, as well as show improved performances when even small amounts of GeoDE (1000 - 2000 images per region) are added to a training dataset. We release the full dataset and code at https://geodiverse-data-collection.cs.princeton.edu/

Telepresence Video Quality Assessment

Video conferencing, which includes both video and audio content, has contributed to dramatic increases in Internet traffic, as the COVID-19 pandemic forced millions of people to work and learn from home. Global Internet traffic of video conferencing has dramatically increased Because of this, efficient and accurate video quality tools are needed to monitor and perceptually optimize telepresence traffic streamed via Zoom, Webex, Meet, etc. However, existing models are limited in their prediction capabilities on multi-modal, live streaming telepresence content. Here we address the significant challenges of Telepresence Video Quality Assessment (TVQA) in several ways. First, we mitigated the dearth of subjectively labeled data by collecting ~2k telepresence videos from different countries, on which we crowdsourced ~80k subjective quality labels. Using this new resource, we created a first-of-a-kind online video quality prediction framework for live streaming, using a multi-modal learning framework with separate pathways to compute visual and audio quality predictions. Our all-in-one model is able to provide accurate quality predictions at the patch, frame, clip, and audiovisual levels. Our model achieves state-of-the-art performance on both existing quality databases and our new TVQA database, at a considerably lower computational expense, making it an attractive solution for mobile and embedded systems.

Making Heads or Tails: Towards Semantically Consistent Visual Counterfactuals

A visual counterfactual explanation replaces image regions in a query image with regions from a distractor image such that the system's decision on the transformed image changes to the distractor class. In this work, we present a novel framework for computing visual counterfactual explanations based on two key ideas. First, we enforce that the \textit{replaced} and \textit{replacer} regions contain the same semantic part, resulting in more semantically consistent explanations. Second, we use multiple distractor images in a computationally efficient way and obtain more discriminative explanations with fewer region replacements. Our approach is $\mathbf{27\%}$ more semantically consistent and an order of magnitude faster than a competing method on three fine-grained image recognition datasets. We highlight the utility of our counterfactuals over existing works through machine teaching experiments where we teach humans to classify different bird species. We also complement our explanations with the vocabulary of parts and attributes that contributed the most to the system's decision. In this task as well, we obtain state-of-the-art results when using our counterfactual explanations relative to existing works, reinforcing the importance of semantically consistent explanations.

Generic Event Boundary Detection: A Benchmark for Event Segmentation

This paper presents a novel task together with a new benchmark for detecting generic, taxonomy-free event boundaries that segment a whole video into chunks. Conventional work in temporal video segmentation and action detection focuses on localizing pre-defined action categories and thus does not scale to generic videos. Cognitive Science has known since last century that humans consistently segment videos into meaningful temporal chunks. This segmentation happens naturally, with no pre-defined event categories and without being explicitly asked to do so. Here, we repeat these cognitive experiments on mainstream CV datasets; with our novel annotation guideline which addresses the complexities of taxonomy-free event boundary annotation, we introduce the task of Generic Event Boundary Detection (GEBD) and the new benchmark Kinetics-GEBD. Through experiment and human study we demonstrate the value of the annotations. We view this as an important stepping stone towards understanding the video as a whole, and believe it has been previously neglected due to a lack of proper task definition and annotations. Further, inspired by the cognitive finding that humans mark boundaries at points where they are unable to predict the future accurately, we explore un-supervised approaches based on temporal predictability. We identify and extensively explore important design factors for GEBD models on the TAPOS dataset and our Kinetics-GEBD while achieving competitive performance and suggesting future work. We will release our annotations and code at CVPR'21 LOVEU Challenge: https://sites.google.com/view/loveucvpr21

Patch-VQ: 'Patching Up' the Video Quality Problem

No-reference (NR) perceptual video quality assessment (VQA) is a complex, unsolved, and important problem to social and streaming media applications. Efficient and accurate video quality predictors are needed to monitor and guide the processing of billions of shared, often imperfect, user-generated content (UGC). Unfortunately, current NR models are limited in their prediction capabilities on real-world, "in-the-wild" UGC video data. To advance progress on this problem, we created the largest (by far) subjective video quality dataset, containing 39, 000 realworld distorted videos and 117, 000 space-time localized video patches ('v-patches'), and 5.5M human perceptual quality annotations. Using this, we created two unique NR-VQA models: (a) a local-to-global region-based NR VQA architecture (called PVQ) that learns to predict global video quality and achieves state-of-the-art performance on 3 UGC datasets, and (b) a first-of-a-kind space-time video quality mapping engine (called PVQ Mapper) that helps localize and visualize perceptual distortions in space and time. We will make the new database and prediction models available immediately following the review process.