Yunsung Lee

Hyojun Go, JinYoung Kim, Yunsung Lee et al. · cmu

Diffusion-based generative models have achieved remarkable success in various domains. It trains a shared model on denoising tasks that encompass different noise levels simultaneously, representing a form of multi-task learning (MTL). However, analyzing and improving diffusion models from an MTL perspective remains under-explored. In particular, MTL can sometimes lead to the well-known phenomenon of negative transfer, which results in the performance degradation of certain tasks due to conflicts between tasks. In this paper, we first aim to analyze diffusion training from an MTL standpoint, presenting two key observations: (O1) the task affinity between denoising tasks diminishes as the gap between noise levels widens, and (O2) negative transfer can arise even in diffusion training. Building upon these observations, we aim to enhance diffusion training by mitigating negative transfer. To achieve this, we propose leveraging existing MTL methods, but the presence of a huge number of denoising tasks makes this computationally expensive to calculate the necessary per-task loss or gradient. To address this challenge, we propose clustering the denoising tasks into small task clusters and applying MTL methods to them. Specifically, based on (O2), we employ interval clustering to enforce temporal proximity among denoising tasks within clusters. We show that interval clustering can be solved using dynamic programming, utilizing signal-to-noise ratio, timestep, and task affinity for clustering objectives. Through this, our approach addresses the issue of negative transfer in diffusion models by allowing for efficient computation of MTL methods. We validate the efficacy of proposed clustering and its integration with MTL methods through various experiments, demonstrating 1) improved generation quality and 2) faster training convergence of diffusion models.

Hyojun Go, Yunsung Lee, Jin-Young Kim et al.

Diffusion-based generative models have achieved remarkable success in image generation. Their guidance formulation allows an external model to plug-and-play control the generation process for various tasks without finetuning the diffusion model. However, the direct use of publicly available off-the-shelf models for guidance fails due to their poor performance on noisy inputs. For that, the existing practice is to fine-tune the guidance models with labeled data corrupted with noises. In this paper, we argue that this practice has limitations in two aspects: (1) performing on inputs with extremely various noises is too hard for a single guidance model; (2) collecting labeled datasets hinders scaling up for various tasks. To tackle the limitations, we propose a novel strategy that leverages multiple experts where each expert is specialized in a particular noise range and guides the reverse process of the diffusion at its corresponding timesteps. However, as it is infeasible to manage multiple networks and utilize labeled data, we present a practical guidance framework termed Practical Plug-And-Play (PPAP), which leverages parameter-efficient fine-tuning and data-free knowledge transfer. We exhaustively conduct ImageNet class conditional generation experiments to show that our method can successfully guide diffusion with small trainable parameters and no labeled data. Finally, we show that image classifiers, depth estimators, and semantic segmentation models can guide publicly available GLIDE through our framework in a plug-and-play manner. Our code is available at https://github.com/riiid/PPAP.

Suhwan Choi, Yunsung Lee, Yubeen Park et al.

Vision Language Action VLA models are typically evaluated using per benchmark scripts maintained independently by each model repository, leading to duplicated code, dependency conflicts, and underspecified protocols. We present vla eval, an open source evaluation harness that decouples model inference from benchmark execution through a WebSocket msgpack protocol with Docker based environment isolation. Models integrate once by implementing a single predict() method; benchmarks integrate once via a four method interface; the full cross evaluation matrix works automatically. A complete evaluation requires only two commands: vla eval serve and vla eval run. The framework supports 13 simulation benchmarks and six model servers. Parallel evaluation via episode sharding and batch inference achieves a 47x throughput improvement, completing 2000 LIBERO episodes in about 18 minutes. Using this infrastructure, we conduct a reproducibility audit of a published VLA model across three benchmarks, finding that all three closely reproduce published values while uncovering undocumented requirements ambiguous termination semantics and hidden normalization statistics that can silently distort results. We additionally release a VLA leaderboard aggregating 657 published results across 17 benchmarks. Framework, evaluation configs, and all reproduction results are publicly available.

Yunsung Lee, Gyuseong Lee, Kwangrok Ryoo et al. · nvidia, utoronto

There are two de facto standard architectures in recent computer vision: Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs). Strong inductive biases of convolutions help the model learn sample effectively, but such strong biases also limit the upper bound of CNNs when sufficient data are available. On the contrary, ViT is inferior to CNNs for small data but superior for sufficient data. Recent approaches attempt to combine the strengths of these two architectures. However, we show these approaches overlook that the optimal inductive bias also changes according to the target data scale changes by comparing various models' accuracy on subsets of sampled ImageNet at different ratios. In addition, through Fourier analysis of feature maps, the model's response patterns according to signal frequency changes, we observe which inductive bias is advantageous for each data scale. The more convolution-like inductive bias is included in the model, the smaller the data scale is required where the ViT-like model outperforms the ResNet performance. To obtain a model with flexible inductive bias on the data scale, we show reparameterization can interpolate inductive bias between convolution and self-attention. By adjusting the number of epochs the model stays in the convolution, we show that reparameterization from convolution to self-attention interpolates the Fourier analysis pattern between CNNs and ViTs. Adapting these findings, we propose Progressive Reparameterization Scheduling (PRS), in which reparameterization adjusts the required amount of convolution-like or self-attention-like inductive bias per layer. For small-scale datasets, our PRS performs reparameterization from convolution to self-attention linearly faster at the late stage layer. PRS outperformed previous studies on the small-scale dataset, e.g., CIFAR-100.

Yunsung Lee, Jin-Young Kim, Hyojun Go et al.

In this paper, we address the performance degradation of efficient diffusion models by introducing Multi-architecturE Multi-Expert diffusion models (MEME). We identify the need for tailored operations at different time-steps in diffusion processes and leverage this insight to create compact yet high-performing models. MEME assigns distinct architectures to different time-step intervals, balancing convolution and self-attention operations based on observed frequency characteristics. We also introduce a soft interval assignment strategy for comprehensive training. Empirically, MEME operates 3.3 times faster than baselines while improving image generation quality (FID scores) by 0.62 (FFHQ) and 0.37 (CelebA). Though we validate the effectiveness of assigning more optimal architecture per time-step, where efficient models outperform the larger models, we argue that MEME opens a new design choice for diffusion models that can be easily applied in other scenarios, such as large multi-expert models.

Jin-Young Kim, Soonwoo Kwon, Hyojun Go et al.

Self-supervised contrastive learning (CL) has achieved state-of-the-art performance in representation learning by minimizing the distance between positive pairs while maximizing that of negative ones. Recently, it has been verified that the model learns better representation with diversely augmented positive pairs because they enable the model to be more view-invariant. However, only a few studies on CL have considered the difference between augmented views, and have not gone beyond the hand-crafted findings. In this paper, we first observe that the score-matching function can measure how much data has changed from the original through augmentation. With the observed property, every pair in CL can be weighted adaptively by the difference of score values, resulting in boosting the performance of the existing CL method. We show the generality of our method, referred to as ScoreCL, by consistently improving various CL methods, SimCLR, SimSiam, W-MSE, and VICReg, up to 3%p in k-NN evaluation on CIFAR-10, CIFAR-100, and ImageNet-100. Moreover, we have conducted exhaustive experiments and ablations, including results on diverse downstream tasks, comparison with possible baselines, and improvement when used with other proposed augmentation methods. We hope our exploration will inspire more research in exploiting the score matching for CL.

Jisu Nam, Yicong Hong, Chun-Hao Paul Huang et al.

Recent advances in video diffusion transformers have enabled interactive gaming world models that allow users to explore generated environments over extended horizons. However, existing approaches struggle with precise action control and long-horizon 3D consistency. Most prior works treat user actions as abstract conditioning signals, overlooking the fundamental geometric coupling between actions and the 3D world, whereby actions induce relative camera motions that accumulate into a global camera pose within a 3D world. In this paper, we establish camera pose as a unifying geometric representation to jointly ground immediate action control and long-term 3D consistency. First, we define a physics-based continuous action space and represent user inputs in the Lie algebra to derive precise 6-DoF camera poses, which are injected into the generative model via a camera embedder to ensure accurate action alignment. Second, we use global camera poses as spatial indices to retrieve relevant past observations, enabling geometrically consistent revisiting of locations during long-horizon navigation. To support this research, we introduce a large-scale dataset comprising 3,000 minutes of authentic human gameplay annotated with camera trajectories and textual descriptions. Extensive experiments show that our approach substantially outperforms state-of-the-art interactive gaming world models in action controllability, long-horizon visual quality, and 3D spatial consistency.

Haebin Seong, Sungmin Kim, Yongjun Cho et al.

While current navigation benchmarks prioritize task success in simplified settings, they neglect the multidimensional economic constraints essential for the real-world commercialization of autonomous delivery systems. We introduce CostNav, an Economic Navigation Benchmark that evaluates physical AI agents through comprehensive economic cost-revenue analysis aligned with real-world business operations. By integrating industry-standard data - such as SEC filings and AIS injury reports - with Isaac Sim's detailed collision and cargo dynamics, CostNav transcends simple task completion to accurately evaluate business value in complex, real-world scenarios. To our knowledge, CostNav is the first work to quantitatively expose the gap between navigation research metrics and commercial viability, revealing that optimizing for task success on a simplified task fundamentally differs from optimizing for real-world economic deployment. Our evaluation of rule-based Nav2 navigation shows that current approaches are not economically viable: the contribution margin is -22.81/run (AMCL) and -12.87/run (GPS), resulting in no break-even point. We challenge the community to develop navigation policies that achieve economic viability on CostNav. We remain method-agnostic, evaluating success solely on the metric of cost rather than the underlying architecture. All resources are available at https://github.com/worv-ai/CostNav.

Junbum Cha, Sanghyuk Chun, Kyungjae Lee et al.

Domain generalization (DG) methods aim to achieve generalizability to an unseen target domain by using only training data from the source domains. Although a variety of DG methods have been proposed, a recent study shows that under a fair evaluation protocol, called DomainBed, the simple empirical risk minimization (ERM) approach works comparable to or even outperforms previous methods. Unfortunately, simply solving ERM on a complex, non-convex loss function can easily lead to sub-optimal generalizability by seeking sharp minima. In this paper, we theoretically show that finding flat minima results in a smaller domain generalization gap. We also propose a simple yet effective method, named Stochastic Weight Averaging Densely (SWAD), to find flat minima. SWAD finds flatter minima and suffers less from overfitting than does the vanilla SWA by a dense and overfit-aware stochastic weight sampling strategy. SWAD shows state-of-the-art performances on five DG benchmarks, namely PACS, VLCS, OfficeHome, TerraIncognita, and DomainNet, with consistent and large margins of +1.6% averagely on out-of-domain accuracy. We also compare SWAD with conventional generalization methods, such as data augmentation and consistency regularization methods, to verify that the remarkable performance improvements are originated from by seeking flat minima, not from better in-domain generalizability. Last but not least, SWAD is readily adaptable to existing DG methods without modification; the combination of SWAD and an existing DG method further improves DG performances. Source code is available at https://github.com/khanrc/swad.

Seokeon Choi, Junhyun Lee, Yunsung Lee et al.

We propose an improved discriminative model prediction method for robust long-term tracking based on a pre-trained short-term tracker. The baseline pre-trained short-term tracker is SuperDiMP which combines the bounding-box regressor of PrDiMP with the standard DiMP classifier. Our tracker RLT-DiMP improves SuperDiMP in the following three aspects: (1) Uncertainty reduction using random erasing: To make our model robust, we exploit an agreement from multiple images after erasing random small rectangular areas as a certainty. And then, we correct the tracking state of our model accordingly. (2) Random search with spatio-temporal constraints: we propose a robust random search method with a score penalty applied to prevent the problem of sudden detection at a distance. (3) Background augmentation for more discriminative feature learning: We augment various backgrounds that are not included in the search area to train a more robust model in the background clutter. In experiments on the VOT-LT2020 benchmark dataset, the proposed method achieves comparable performance to the state-of-the-art long-term trackers. The source code is available at: https://github.com/bismex/RLT-DIMP.

Suwhan Choi, Jaeyoon Jung, Haebin Seong et al.

Large language models leverage internet-scale text data, yet embodied AI remains constrained by the prohibitive costs of physical trajectory collection. Desktop environments -- particularly gaming -- offer a compelling alternative: they provide rich sensorimotor interactions at scale while maintaining the structured observation-action coupling essential for embodied learning. We present D2E (Desktop to Embodied AI), a framework that demonstrates desktop interactions can serve as an effective pretraining substrate for robotics embodied AI tasks. Unlike prior work that remained domain-specific (e.g., VPT for Minecraft) or kept data proprietary (e.g., SIMA), D2E establishes a complete pipeline from scalable desktop data collection to verified transfer in embodied domains. Our framework comprises three components: (1) the OWA Toolkit that unifies diverse desktop interactions into a standardized format with 152x compression, (2) the Generalist-IDM that achieves strong zero-shot generalization across unseen games through timestamp-based event prediction, enabling internet-scale pseudo-labeling, and (3) VAPT that transfers desktop-pretrained representations to physical manipulation and navigation. Using 1.3K+ hours of data (259 hours of human demonstrations, and 1K+ hours of pseudo-labeled gameplay), we achieve a total of 96.6% success rate on LIBERO manipulation and 83.3% on CANVAS navigation benchmarks. This validates that sensorimotor primitives in digital interactions exhibit sufficient invariance to transfer meaningfully to physical embodied tasks, establishing desktop pretraining as a practical paradigm for robotics. We will make all our work public, including the OWA toolkit, datasets of human-collected and pseudo-labeled, and VAPT-trained models available at https://worv-ai.github.io/d2e/

Hyun Seung Lee, Seungtaek Choi, Yunsung Lee et al.

Text classification in education, usually called auto-tagging, is the automated process of assigning relevant tags to educational content, such as questions and textbooks. However, auto-tagging suffers from a data scarcity problem, which stems from two major challenges: 1) it possesses a large tag space and 2) it is multi-label. Though a retrieval approach is reportedly good at low-resource scenarios, there have been fewer efforts to directly address the data scarcity problem. To mitigate these issues, here we propose a novel retrieval approach CEAA that provides effective learning in educational text classification. Our main contributions are as follows: 1) we leverage transfer learning from question-answering datasets, and 2) we propose a simple but effective data augmentation method introducing cross-encoder style texts to a bi-encoder architecture for more efficient inference. An extensive set of experiments shows that our proposed method is effective in multi-label scenarios and low-resource tags compared to state-of-the-art models.

Shinhyeok Oh, Hyojun Go, Hyeongdon Moon et al.

Question generation (QG) is the task of generating a valid and fluent question based on a given context and the target answer. According to various purposes, even given the same context, instructors can ask questions about different concepts, and even the same concept can be written in different ways. However, the evaluation for QG usually depends on single reference-based similarity metrics, such as n-gram-based metric or learned metric, which is not sufficient to fully evaluate the potential of QG methods. To this end, we propose to paraphrase the reference question for a more robust QG evaluation. Using large language models such as GPT-3, we created semantically and syntactically diverse questions, then adopt the simple aggregation of the popular evaluation metrics as the final scores. Through our experiments, we found that using multiple (pseudo) references is more effective for QG evaluation while showing a higher correlation with human evaluations than evaluation with a single reference.

Yunsung Lee, Teakgyu Hong, Han-Cheol Cho et al.

Recently, self-supervised methods show remarkable achievements in image-level representation learning. Nevertheless, their image-level self-supervisions lead the learned representation to sub-optimal for dense prediction tasks, such as object detection, instance segmentation, etc. To tackle this issue, several recent self-supervised learning methods have extended image-level single embedding to pixel-level dense embeddings. Unlike image-level representation learning, due to the spatial deformation of augmentation, it is difficult to sample pixel-level positive pairs. Previous studies have sampled pixel-level positive pairs using the winner-takes-all among similarity or thresholding warped distance between dense embeddings. However, these naive methods can be struggled by background clutter and outliers problems. In this paper, we introduce Hough Contrastive Learning (HoughCL), a Hough space based method that enforces geometric consistency between two dense features. HoughCL achieves robustness against background clutter and outliers. Furthermore, compared to baseline, our dense positive pairing method has no additional learnable parameters and has a small extra computation cost. Compared to previous works, our method shows better or comparable performance on dense prediction fine-tuning tasks.

Seokju Cho, Sunghwan Hong, Sangryul Jeon et al.

We propose a novel cost aggregation network, called Cost Aggregation Transformers (CATs), to find dense correspondences between semantically similar images with additional challenges posed by large intra-class appearance and geometric variations. Cost aggregation is a highly important process in matching tasks, which the matching accuracy depends on the quality of its output. Compared to hand-crafted or CNN-based methods addressing the cost aggregation, in that either lacks robustness to severe deformations or inherit the limitation of CNNs that fail to discriminate incorrect matches due to limited receptive fields, CATs explore global consensus among initial correlation map with the help of some architectural designs that allow us to fully leverage self-attention mechanism. Specifically, we include appearance affinity modeling to aid the cost aggregation process in order to disambiguate the noisy initial correlation maps and propose multi-level aggregation to efficiently capture different semantics from hierarchical feature representations. We then combine with swapping self-attention technique and residual connections not only to enforce consistent matching but also to ease the learning process, which we find that these result in an apparent performance boost. We conduct experiments to demonstrate the effectiveness of the proposed model over the latest methods and provide extensive ablation studies. Project page is available at : https://sunghwanhong.github.io/CATs/.

Junsoo Lee, Eungyeup Kim, Yunsung Lee et al.

This paper tackles the automatic colorization task of a sketch image given an already-colored reference image. Colorizing a sketch image is in high demand in comics, animation, and other content creation applications, but it suffers from information scarcity of a sketch image. To address this, a reference image can render the colorization process in a reliable and user-driven manner. However, it is difficult to prepare for a training data set that has a sufficient amount of semantically meaningful pairs of images as well as the ground truth for a colored image reflecting a given reference (e.g., coloring a sketch of an originally blue car given a reference green car). To tackle this challenge, we propose to utilize the identical image with geometric distortion as a virtual reference, which makes it possible to secure the ground truth for a colored output image. Furthermore, it naturally provides the ground truth for dense semantic correspondence, which we utilize in our internal attention mechanism for color transfer from reference to sketch input. We demonstrate the effectiveness of our approach in various types of sketch image colorization via quantitative as well as qualitative evaluation against existing methods.