Selected work

Representative Papers

PixReenact: Pixel-Conditioned Causal Video Diffusion for Streaming Head-Avatar Reenactment

Oct 04, 2026

This study addresses the limitations of streaming head reenactment, including reliance on specialized representations, high latency from offline generation, and long-term identity drift. To overcome these challenges, this work proposes a pixel-conditioned causal video diffusion framework that directly utilizes VAE-encoded frames as driving conditions. The method incorporates cross-identity pseudo-supervision, a self-rolling anti-drift mechanism, and state-aware dual-teacher distillation to achieve low-latency real-time generation while preserving long-sequence identity consistency. Experimental results demonstrate that the proposed model exhibits strong robustness under extreme viewpoints, maintains stable identity over extended sequences, and achieves an average time-to-first-frame latency of only 239 milliseconds.

0 citationsRead paper

From A to B to A: Palindromic Zero-Shot Voice Conversion with Non-Parallel Data

Jun 07, 2026

This work proposes a zero-shot cross-lingual voice conversion method that operates without parallel speech data. To address the challenges posed by the absence of explicit alignments and multilingual training corpora, the approach leverages WavLM speech representations and constructs synthetic training pairs from non-parallel source and target utterances via k-nearest neighbor retrieval, employing a “synthetic-to-real” supervised learning paradigm. Additionally, it incorporates a speaker loss derived from a pretrained speaker verification model to enhance target speaker consistency. Trained exclusively on English data, the method achieves high naturalness and strong speaker similarity across multiple languages, significantly outperforming existing baselines and demonstrating robust cross-lingual transfer capability and practical applicability.

0 citationsRead paper

VidMsg: A Benchmark for Implicit Message Inference in Short Videos

Jun 02, 2026

This work addresses the challenge that existing video understanding methods struggle to capture the implicit intent of creators in short videos. To this end, we introduce VidMsg, the first benchmark for short-video understanding centered on latent messages. Built through a message-first pipeline, VidMsg comprises 400 YouTube videos spanning nine broad themes and 52 fine-grained message categories. By leveraging large language models to generate indirectly expressed scenarios and incorporating human-curated non-explicit videos, the benchmark supports bidirectional message–video retrieval and multiple-choice question answering. We also propose VidVec-Msg as a baseline method. Experimental results reveal that state-of-the-art models perform substantially below human levels, highlighting the critical need for pragmatic reasoning and fine-grained semantic discrimination, and thereby filling a key gap in evaluating implicit intent understanding in video content.

0 citationsRead paper

VidVec: Unlocking Video MLLM Embeddings for Video-Text Retrieval

Feb 08, 2026

This work addresses the underperformance of existing multimodal large language model (MLLM)-based video embedding approaches compared to specialized video foundation models in video–text retrieval tasks. Through a systematic analysis of intermediate-layer features in MLLMs, the authors reveal that rich video semantics are already embedded within these representations. They propose a lightweight, vision-supervision-free training paradigm that leverages only text-summary alignment for embedding learning, combined with a calibrated MLLM head to enable zero-shot retrieval. The method significantly outperforms current state-of-the-art approaches across multiple standard video retrieval benchmarks, without requiring any visual-domain fine-tuning. These results demonstrate the effectiveness and superiority of purely text-alignment-driven video embedding learning.

0 citationsRead paper

Dicta-LM 3.0: Advancing The Frontier of Hebrew Sovereign LLMs

Feb 02, 2026

This work addresses the scarcity of high-quality open-source large language models (LLMs) for low-resource languages like Hebrew, which hinders localized applications. We present the first systematic effort to develop a family of Hebrew LLMs at three scales—1.7B, 12B, and 24B parameters—based on Mistral-Small-3.1, NVIDIA Nemotron Nano V2, and Qwen3-1.7B, respectively. These models are adapted using large-scale Hebrew–English mixed corpora, support 65K-token context lengths and tool calling, and are released in both base and chat variants. Additionally, we introduce the first comprehensive evaluation benchmark for Hebrew chat models, demonstrating strong performance across tasks including translation, summarization, Winograd schema resolution, Israeli commonsense question answering, and nikud (vowel diacritic) restoration. The proposed framework is readily generalizable to other non-English languages and significantly advances Hebrew natural language processing.

0 citationsRead paper

Recent publications

Latest Papers

PixReenact: Pixel-Conditioned Causal Video Diffusion for Streaming Head-Avatar Reenactment

Oct 04, 2026

This study addresses the limitations of streaming head reenactment, including reliance on specialized representations, high latency from offline generation, and long-term identity drift. To overcome these challenges, this work proposes a pixel-conditioned causal video diffusion framework that directly utilizes VAE-encoded frames as driving conditions. The method incorporates cross-identity pseudo-supervision, a self-rolling anti-drift mechanism, and state-aware dual-teacher distillation to achieve low-latency real-time generation while preserving long-sequence identity consistency. Experimental results demonstrate that the proposed model exhibits strong robustness under extreme viewpoints, maintains stable identity over extended sequences, and achieves an average time-to-first-frame latency of only 239 milliseconds.

0 citationsRead paper

From A to B to A: Palindromic Zero-Shot Voice Conversion with Non-Parallel Data

Jun 07, 2026

This work proposes a zero-shot cross-lingual voice conversion method that operates without parallel speech data. To address the challenges posed by the absence of explicit alignments and multilingual training corpora, the approach leverages WavLM speech representations and constructs synthetic training pairs from non-parallel source and target utterances via k-nearest neighbor retrieval, employing a “synthetic-to-real” supervised learning paradigm. Additionally, it incorporates a speaker loss derived from a pretrained speaker verification model to enhance target speaker consistency. Trained exclusively on English data, the method achieves high naturalness and strong speaker similarity across multiple languages, significantly outperforming existing baselines and demonstrating robust cross-lingual transfer capability and practical applicability.

0 citationsRead paper

VidMsg: A Benchmark for Implicit Message Inference in Short Videos

Jun 02, 2026

This work addresses the challenge that existing video understanding methods struggle to capture the implicit intent of creators in short videos. To this end, we introduce VidMsg, the first benchmark for short-video understanding centered on latent messages. Built through a message-first pipeline, VidMsg comprises 400 YouTube videos spanning nine broad themes and 52 fine-grained message categories. By leveraging large language models to generate indirectly expressed scenarios and incorporating human-curated non-explicit videos, the benchmark supports bidirectional message–video retrieval and multiple-choice question answering. We also propose VidVec-Msg as a baseline method. Experimental results reveal that state-of-the-art models perform substantially below human levels, highlighting the critical need for pragmatic reasoning and fine-grained semantic discrimination, and thereby filling a key gap in evaluating implicit intent understanding in video content.

0 citationsRead paper

VidVec: Unlocking Video MLLM Embeddings for Video-Text Retrieval

Feb 08, 2026

This work addresses the underperformance of existing multimodal large language model (MLLM)-based video embedding approaches compared to specialized video foundation models in video–text retrieval tasks. Through a systematic analysis of intermediate-layer features in MLLMs, the authors reveal that rich video semantics are already embedded within these representations. They propose a lightweight, vision-supervision-free training paradigm that leverages only text-summary alignment for embedding learning, combined with a calibrated MLLM head to enable zero-shot retrieval. The method significantly outperforms current state-of-the-art approaches across multiple standard video retrieval benchmarks, without requiring any visual-domain fine-tuning. These results demonstrate the effectiveness and superiority of purely text-alignment-driven video embedding learning.

0 citationsRead paper

Dicta-LM 3.0: Advancing The Frontier of Hebrew Sovereign LLMs

Feb 02, 2026

This work addresses the scarcity of high-quality open-source large language models (LLMs) for low-resource languages like Hebrew, which hinders localized applications. We present the first systematic effort to develop a family of Hebrew LLMs at three scales—1.7B, 12B, and 24B parameters—based on Mistral-Small-3.1, NVIDIA Nemotron Nano V2, and Qwen3-1.7B, respectively. These models are adapted using large-scale Hebrew–English mixed corpora, support 65K-token context lengths and tool calling, and are released in both base and chat variants. Additionally, we introduce the first comprehensive evaluation benchmark for Hebrew chat models, demonstrating strong performance across tasks including translation, summarization, Winograd schema resolution, Israeli commonsense question answering, and nikud (vowel diacritic) restoration. The proposed framework is readily generalizable to other non-English languages and significantly advances Hebrew natural language processing.

0 citationsRead paper