🏢 Peking University

HermesFlow seamlessly bridges the understanding-generation gap in MLLMs using a novel Pair-DPO framework and self-play optimization on homologous data, achieving significant performance improvements.

Diffusion-Sharpening enhances diffusion model fine-tuning by optimizing sampling trajectories, achieving faster convergence and high inference efficiency without extra NFEs, leading to improved alignm…

HealthGPT: A novel medical vision-language model unifying comprehension and generation via heterogeneous knowledge adaptation, achieving superior performance on various medical tasks.

Next-Block Prediction (NBP) revolutionizes video generation by using a semi-autoregressive model that predicts blocks of video content simultaneously, resulting in significantly faster inference.

Magic141 generates one-minute video clips in under a minute by cleverly factorizing the generation task and employing optimization techniques.

InferenceGuard ensures almost-sure safe LLM responses at inference time by framing safe generation as a constrained Markov Decision Process in the LLM’s latent space, achieving high safety rates witho…

DIFFSPLAT repurposes 2D image diffusion models to natively generate high-quality 3D Gaussian splats, overcoming limitations in existing 3D generation methods.

ARWKV: A novel RNN-attention-based language model, distilled from a larger model, achieves strong performance using significantly fewer resources, opening a new path in efficient language model develo…

Researchers significantly enhanced autoregressive image generation by integrating chain-of-thought reasoning strategies, achieving a remarkable +24% improvement on the GenEval benchmark.

PaSa: An LLM agent autonomously performs comprehensive academic paper searches, outperforming existing methods by efficiently combining search tools, paper reading, and citation analysis, optimized vi…

Sa2VA marries SAM2 and LLaVA for dense grounded image and video understanding, achieving state-of-the-art results on multiple benchmarks.

Friends-MMC: A new dataset facilitates multi-modal multi-party conversation understanding by providing 24,000+ utterances with video, audio, and speaker annotations, enabling advancements in character…

ROBUSTFT tackles noisy data in LLM fine-tuning by using multi-expert noise detection and context-enhanced relabeling, significantly boosting model performance in noisy scenarios.

Boosting autoregressive visual generation speed by 3.6-9.5x, this research introduces parallel processing while preserving model simplicity and generation quality.

Boosting code generation accuracy, Outcome-Refining Process Supervision (ORPS) uses execution feedback and structured reasoning to refine code, achieving significant improvements across models and dat…

MOVIS enhances 3D scene generation by improving cross-view consistency in multi-object novel view synthesis.

BrushEdit revolutionizes interactive image editing with instructions & inpainting.

DisPose disentangles pose guidance for controllable human image animation, generating diverse animations while preserving appearance consistency using only sparse skeleton pose input, eliminating the …

DiffSensei: A new framework generates customized manga with dynamic multi-character control using multi-modal LLMs and diffusion models, outperforming existing methods.

HumanEdit: A new human-rewarded dataset revolutionizes instruction-based image editing by providing high-quality, diverse image pairs with detailed instructions, enabling precise model evaluation and …