Junwan Kim

Hyeongmin Lee, Jin-Young Kim, Kyungjune Baek et al.

In this work, we discuss evaluating video foundation models in a fair and robust manner. Unlike language or image foundation models, many video foundation models are evaluated with differing parameters (such as sampling rate, number of frames, pretraining steps, etc.), making fair and robust comparisons challenging. Therefore, we present a carefully designed evaluation framework for measuring two core capabilities of video comprehension: appearance and motion understanding. Our findings reveal that existing video foundation models, whether text-supervised like UMT or InternVideo2, or self-supervised like V-JEPA, exhibit limitations in at least one of these capabilities. As an alternative, we introduce TWLV-I, a new video foundation model that constructs robust visual representations for both motion- and appearance-based videos. Based on the average top-1 accuracy of linear probing on five action recognition benchmarks, pretrained only on publicly accessible datasets, our model shows a 4.6%p improvement compared to V-JEPA (ViT-L) and a 7.7%p improvement compared to UMT (ViT-L). Even when compared to much larger models, our model demonstrates a 7.2%p improvement compared to DFN (ViT-H), a 2.7%p improvement compared to V-JEPA (ViT-H) and a 2.8%p improvement compared to InternVideo2 (ViT-g). We provide embedding vectors obtained by TWLV-I from videos of several commonly used video benchmarks, along with evaluation source code that can directly utilize these embeddings. The code is available at https://github.com/twelvelabs-io/video-embeddings-evaluation-framework.

Junwan Kim, Hyunkyung Bae

Multimodal large language models (MLLMs) have recently demonstrated strong capabilities in understanding and generating responses from diverse visual inputs, including high-resolution images and long video sequences. As these models scale to richer visual representations, inference increasingly relies on storing large numbers of vision tokens in the key-value (KV) cache, making memory consumption a central bottleneck. Existing methods address this issue by identifying redundancy in vision tokens and compressing the cache, but such compression is typically applied only after all inputs are processed, resulting in high peak memory usage during the prefill stage. In this work, we show that MLLMs exhibit inherent structural regularities and representational redundancy that can be exploited to control memory growth throughout inference. Based on this insight, we propose a sequential input-compression mechanism that enforces a fixed memory budget by performing structure-aware key-value cache compression during the prefill process. This approach substantially reduces peak memory usage while maintaining generative performance with only minimal degradation, enabling more practical and memory-efficient multimodal inference.

Chanhyuk Park, Jungbin Cho, Junwan Kim et al.

This work presents an audio-visual interactive chatbot (AVIN-Chat) system that allows users to have face-to-face conversations with 3D avatars in real-time. Compared to the previous chatbot services, which provide text-only or speech-only communications, the proposed AVIN-Chat can offer audio-visual communications providing users with a superior experience quality. In addition, the proposed AVIN-Chat emotionally speaks and expresses according to the user's emotional state. Thus, it enables users to establish a strong bond with the chatbot system, increasing the user's immersion. Through user subjective tests, it is demonstrated that the proposed system provides users with a higher sense of immersion than previous chatbot systems. The demonstration video is available at https://www.youtube.com/watch?v=Z74uIV9k7_k.

Jisoo Kim, Wooseok Seo, Junwan Kim et al.

With growing interest in deploying text-to-video (T2V) models in resource-constrained environments, reducing their high computational cost has become crucial, leading to extensive research on pruning and knowledge distillation methods while maintaining performance. However, existing distillation methods primarily rely on supervised fine-tuning (SFT), which often leads to mode collapse as pruned models with reduced capacity fail to directly match the teacher's outputs, ultimately resulting in degraded quality. To address this challenge, we propose an effective distillation method, ReDPO, that integrates DPO and SFT. Our approach leverages DPO to guide the student model to focus on recovering only the targeted properties, rather than passively imitating the teacher, while also utilizing SFT to enhance overall performance. We additionally propose V.I.P., a novel framework for filtering and curating high-quality pair datasets, along with a step-by-step online approach for calibrated training. We validate our method on two leading T2V models, VideoCrafter2 and AnimateDiff, achieving parameter reduction of 36.2% and 67.5% each, while maintaining or even surpassing the performance of full models. Further experiments demonstrate the effectiveness of both ReDPO and V.I.P. framework in enabling efficient and high-quality video generation. Our code and videos are available at https://jiiiisoo.github.io/VIP.github.io/.

Junwan Kim, Jiho Park, Seonghu Jeon et al.

Flow matching has recently emerged as a promising alternative to diffusion-based generative models, particularly for text-to-image generation. Despite its flexibility in allowing arbitrary source distributions, most existing approaches rely on a standard Gaussian distribution, a choice inherited from diffusion models, and rarely consider the source distribution itself as an optimization target in such settings. In this work, we show that principled design of the source distribution is not only feasible but also beneficial at the scale of modern text-to-image systems. Specifically, we propose learning a condition-dependent source distribution under flow matching objective that better exploit rich conditioning signals. We identify key failure modes that arise when directly incorporating conditioning into the source, including distributional collapse and instability, and show that appropriate variance regularization and directional alignment between source and target are critical for stable and effective learning. We further analyze how the choice of target representation space impacts flow matching with structured sources, revealing regimes in which such designs are most effective. Extensive experiments across multiple text-to-image benchmarks demonstrate consistent and robust improvements, including up to a 3x faster convergence in FID, highlighting the practical benefits of a principled source distribution design for conditional flow matching.

Raehyuk Jung, Hyojun Go, Jaehyuk Yi et al.

This technical report introduces Pegasus-1, a multimodal language model specialized in video content understanding and interaction through natural language. Pegasus-1 is designed to address the unique challenges posed by video data, such as interpreting spatiotemporal information, to offer nuanced video content comprehension across various lengths. This technical report overviews Pegasus-1's architecture, training strategies, and its performance in benchmarks on video conversation, zero-shot video question answering, and video summarization. We also explore qualitative characteristics of Pegasus-1 , demonstrating its capabilities as well as its limitations, in order to provide readers a balanced view of its current state and its future direction.

Jungbin Cho, Junwan Kim, Jisoo Kim et al.

Human motion is inherently continuous and dynamic, posing significant challenges for generative models. While discrete generation methods are widely used, they suffer from limited expressiveness and frame-wise noise artifacts. In contrast, continuous approaches produce smoother, more natural motion but often struggle to adhere to conditioning signals due to high-dimensional complexity and limited training data. To resolve this 'discord' between discrete and continuous representations we introduce DisCoRD: Discrete Tokens to Continuous Motion via Rectified Flow Decoding, a novel method that leverages rectified flow to decode discrete motion tokens in the continuous, raw motion space. Our core idea is to frame token decoding as a conditional generation task, ensuring that DisCoRD captures fine-grained dynamics and achieves smoother, more natural motions. Compatible with any discrete-based framework, our method enhances naturalness without compromising faithfulness to the conditioning signals on diverse settings. Extensive evaluations demonstrate that DisCoRD achieves state-of-the-art performance, with FID of 0.032 on HumanML3D and 0.169 on KIT-ML. These results establish DisCoRD as a robust solution for bridging the divide between discrete efficiency and continuous realism. Project website: https://whwjdqls.github.io/discord-motion/

Heeji Yoon, Jaewoo Jung, Junwan Kim et al.

Multimodal large language models (MLLMs) trained with visual instruction tuning have achieved strong performance across diverse tasks, yet they remain limited in vision-centric tasks such as object counting or spatial reasoning. We attribute this gap to the prevailing text-only supervision paradigm, which provides only indirect guidance for the visual pathway and often leads MLLMs to discard fine-grained visual details during training. In this paper, we present VIsual Representation ALignment (VIRAL), a simple yet effective regularization strategy that aligns the internal visual representations of MLLMs with those of pre-trained vision foundation models (VFMs). By explicitly enforcing this alignment, VIRAL enables the model not only to retain critical visual details from the input vision encoder but also to complement additional visual knowledge from VFMs, thereby enhancing its ability to reason over complex visual inputs. Our experiments demonstrate consistent improvements across all tasks on widely adopted multimodal benchmarks. Furthermore, we conduct comprehensive ablation studies to validate the key design choices underlying our framework. We believe this simple finding opens up an important direction for the effective integration of visual information in training MLLMs.

Sangbeom Lim, Junwan Kim, Heeji Yoon et al.

Region-level captioning aims to generate natural language descriptions for specific image regions while highlighting their distinguishing features. However, existing methods struggle to produce unique captions across multi-granularity, limiting their real-world applicability. To address the need for detailed region-level understanding, we introduce URECA dataset, a large-scale dataset tailored for multi-granularity region captioning. Unlike prior datasets that focus primarily on salient objects, URECA dataset ensures a unique and consistent mapping between regions and captions by incorporating a diverse set of objects, parts, and background elements. Central to this is a stage-wise data curation pipeline, where each stage incrementally refines region selection and caption generation. By leveraging Multimodal Large Language Models (MLLMs) at each stage, our pipeline produces distinctive and contextually grounded captions with improved accuracy and semantic diversity. Building upon this dataset, we present URECA, a novel captioning model designed to effectively encode multi-granularity regions. URECA maintains essential spatial properties such as position and shape through simple yet impactful modifications to existing MLLMs, enabling fine-grained and semantically rich region descriptions. Our approach introduces dynamic mask modeling and a high-resolution mask encoder to enhance caption uniqueness. Experiments show that URECA achieves state-of-the-art performance on URECA dataset and generalizes well to existing region-level captioning benchmarks.