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
Minglei Shi, Haolin Wang, Borui Zhang +11 authors
SVG-T2I framework enables high-quality text-to-image synthesis by training diffusion models within visual foundation model representation space, achieving competitive performance on benchmark datasets.
Guibao Shen, Yihua Du, Wenhang Ge +7 authors
StereoPilot, a feed-forward model leveraging a learnable domain switcher and cycle consistency loss, synthesizes high-quality stereo video directly without depth maps, outperforming existing methods in visual fidelity and computational efficiency.
Team Olmo, Allyson Ettinger, Amanda Bertsch +65 authors
Olmo 3, a family of state-of-the-art fully-open language models at 7B and 32B parameter scales, excels in long-context reasoning, function calling, coding, instruction following, general chat, and knowledge recall.
Enshen Zhou, Cheng Chi, Yibo Li +10 authors
RoboTracer, a 3D-aware visual language model, enhances spatial tracing by combining supervised and reinforcement fine-tuning with a universal spatial encoder and regression-supervised decoder, achieving state-of-the-art performance on TraceSpatial-Bench.
Ahmadreza Jeddi, Hakki Can Karaimer, Hue Nguyen +8 authors
PuzzleCraft presents a supervision-free framework for scaling vision-centric RLVR using lightweight puzzle environments with built-in verification, incorporating an exploration-aware curriculum and a new consistency metric to improve reasoning robustness and downstream performance.
Xin Lin, Meixi Song, Dizhe Zhang +6 authors
A panoramic metric depth foundation model using DINOv3-Large and a three-stage pseudo-label pipeline achieves robust performance across diverse real-world scenes.
Bingxiang He, Zekai Qu, Zeyuan Liu +9 authors
JustRL achieves state-of-the-art performance on reasoning models with minimal complexity, using single-stage training and fixed hyperparameters, outperforming sophisticated approaches in terms of compute and stability.
Kaixin Ding, Yang Zhou, Xi Chen +5 authors
Alchemist, a meta-gradient-based framework, automatically selects high-quality subsets from large-scale text-image datasets to improve visual quality and training efficiency in Text-to-Image models.
Jooyeol Yun, Jaegul Choo
A framework aggregates weak predictions to recover semantic structure, enabling coherent SVG animations and improving VLM interactions with vector graphics.
Ye Fang, Tong Wu, Valentin Deschaintre +6 authors
V-RGBX presents an end-to-end framework for intrinsic-aware video editing that combines video inverse rendering, photorealistic video synthesis, and keyframe-based editing with physically grounded intrinsic channel manipulation.
Lars Mescheder, Wei Dong, Shiwei Li +10 authors
SHARP synthesizes photorealistic views from a single image using a 3D Gaussian representation, achieving state-of-the-art results with rapid processing.
Sihui Ji, Xi Chen, Shuai Yang +3 authors
MemFlow dynamically updates a memory bank by retrieving relevant historical frames for each video chunk, ensuring narrative coherence and generation efficiency with minimal computational overhead.
Hanlin Wang, Hao Ouyang, Qiuyu Wang +12 authors
WorldCanvas generates coherent, controllable world events using a multimodal framework that integrates text, trajectories, and reference images.
Tingyang Chen, Cong Fu, Jiahua Wu +6 authors
Iceberg is a benchmark suite that evaluates vector similarity search methods in real-world contexts, identifying key sources of performance degradation and providing guidance for selecting and tuning these methods.
Jun Zhang, Teng Wang, Yuying Ge +4 authors
TimeLens establishes a robust baseline for video temporal grounding by improving benchmark quality, addressing noisy training data, and developing efficient algorithmic design principles for multimodal large language models.
Giorgos Petsangourakis, Christos Sgouropoulos, Bill Psomas +3 authors
REGLUE, a unified latent diffusion framework, enhances image synthesis by jointly modeling VAE latents, patch-level VFM semantics, and global tokens, improving semantic supervision and convergence.
Mengxi Xiao, Kailai Yang, Pengde Zhao +10 authors
A unified framework and benchmark called MentraSuite is introduced to enhance reliable mental-health reasoning in large language models, with a post-trained model named Mindora achieving superior performance through a hybrid SFT-RL approach optimized for consistency and faithfulness.
Linghui Shen, Mingyue Cui, Xingyi Yang
DeContext defends against unauthorized in-context image editing by weakening cross-attention pathways in multimodal attention layers, preserving visual quality while blocking unwanted modifications.
Zhenwen Liang, Sidi Lu, Wenhao Yu +4 authors
G2RL is a gradient-guided reinforcement learning framework that improves large language model reasoning by aligning exploration with the model's own update geometry rather than using external heuristics.
Jingli Lin, Runsen Xu, Shaohao Zhu +11 authors
MMSI-Video-Bench is a comprehensive benchmark for video-based spatial intelligence in MLLMs, revealing significant gaps between human and AI performance and highlighting challenges in geometric reasoning, motion grounding, and cross-video correspondence.
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