Explore · 每周精选
发现最受关注的研究论文,追踪研究趋势,订阅感兴趣的期刊与关键词。
50 篇论文 · 按点赞排序
Dongjie Yang, Suyuan Huang, Chengqiang Lu +5 authors
Vript introduces a high-quality video-text dataset with detailed captions including camera operations, enhancing video captioning and generation, and proposes Vriptor, a top-performing model with Vript-Hard, a new benchmark for video understanding challenges.
Zhihang Yuan, Pu Lu, Hanling Zhang +6 authors
DiTFastAttn reduces computational complexity in Diffusion Transformers by addressing spatial, temporal, and conditional redundancies.
Yushi Hu, Weijia Shi, Xingyu Fu +5 authors
Sketchpad introduced a visual sketchpad with drawing tools for multimodal LMs, enhancing performance across various math and visual reasoning tasks through human-like sketching and the use of specialist vision models.
Justin Cosentino, Anastasiya Belyaeva, Xin Liu +31 authors
A Personal Health Large Language Model (PH-LLM) fine-tuned from Gemini demonstrates substantial capabilities in processing and interpreting personal health data, offering personalized insights and outperforming experts in certain domain-specific tasks.
Tejaswini Pedapati, Amit Dhurandhar, Soumya Ghosh +2 authors
A framework using engineered features and logistic regression effectively estimates the confidence of large language models with black-box access, outperforming existing methods on multiple datasets.
Dongfu Jiang, Max Ku, Tianle Li +4 authors
GenAI-Arena, an open platform, leverages user feedback to evaluate generative models across text-to-image, text-to-video, and image editing tasks, demonstrating existing multimodal models' shortcomings in assessing generated content quality.
Amil Dravid, Yossi Gandelsman, Kuan-Chieh Wang +4 authors
A large dataset of fine-tuned diffusion models for different visual identities reveals a manifold representing identity in the weight space, enabling sampling, semantic editing, and inversion tasks.
Zhilin Wang, Yi Dong, Olivier Delalleau +6 authors
NVIDIA's HelpSteer2 dataset and SteerLM 2.0 model improve LLM response quality and alignment using permissively licensed preference data.
Xin Jin, Pengyi Jiao, Zheng-Peng Duan +4 authors
LE3D enhances 3D Gaussian Splatting for real-time HDR rendering from RAW images, particularly in low light, by improving structure-from-motion and using a Color MLP for color representation.
Fei Wang, Xingyu Fu, James Y. Huang +18 authors
MuirBench is a benchmark for evaluating robust multi-image understanding in multimodal LLMs, highlighting challenges and the need for improvements in handling multi-image tasks.
Xinzhi Mu, Li Chen, Bohan Chen +5 authors
The introduced FontStudio system uses a shape-adaptive diffusion model to generate text effects for multilingual fonts, maintaining coherence across characters and achieving superior aesthetic results compared to Adobe Firefly.
Tianwen Wei, Bo Zhu, Liang Zhao +13 authors
Skywork-MoE, a high-performance mixture-of-experts language model, uses upcycling and innovative techniques like gating logit normalization and adaptive auxiliary loss coefficients to achieve strong performance across benchmarks.
Juan Manuel Zambrano Chaves, Eric Wang, Tao Tu +7 authors
Tx-LLM, a specialized large language model fine-tuned from PaLM-2, encodes diverse therapeutic knowledge and achieves superior performance across various drug discovery tasks by interleaving chemical data with free-text.
Matthieu Futeral, Armel Zebaze, Pedro Ortiz Suarez +5 authors
mOSCAR, a large-scale multilingual and multimodal document corpus, enhances few-shot learning performance in multilingual image-text tasks when used to train models.
Xiaoshuai Song, Muxi Diao, Guanting Dong +13 authors
CS-Bench evaluates large language models across various subfields of computer science, uncovering performance trends and pinpointing areas for improvement.
Sanyuan Chen, Shujie Liu, Long Zhou +6 authors
VALL-E 2 achieves human parity in zero-shot text-to-speech synthesis through improvements in sampling and code modeling.
Roman Bachmann, Oğuzhan Fatih Kar, David Mizrahi +6 authors
A single multimodal model trained on a diverse set of modalities and tasks through co-learning and discrete tokenization achieves superior performance and flexibility compared to existing models.
Kai Wang, Shijian Deng, Jing Shi +2 authors
AV-DiT, an efficient audio-visual diffusion transformer, generates high-quality joint audio-video content using a shared pretrained DiT backbone with modality-specific lightweight adaptations.
Rishit Dagli, Shivesh Prakash, Robert Wu +1 authors
SEE-2-SOUND generates spatial audio for visual content by decomposing the task into identifying visual regions, locating them in 3D space, generating mono-audio for each, and integrating it into spatial audio, addressing a gap in current audio generation models.
Yucheng Han, Rui Wang, Chi Zhang +4 authors
EMMA enhances multi-modal image generation by integrating text and visual inputs into a diffusion model through a Multi-modal Feature Connector and attention mechanism, allowing for high-quality, context-aware image and video creation.
北京市昌平区探索星信息技术及软件开发工作室
京ICP备2026059466号