🏢 Tencent AI Lab

Hunyuan3D 2.0: A groundbreaking open-source system generating high-resolution, textured 3D assets using scalable diffusion models, exceeding state-of-the-art performance.

XMusic: A new framework generates high-quality, emotionally controllable symbolic music from various prompts (images, videos, text, tags, humming).

CityDreamer4D generates realistic, unbounded 4D city models by cleverly separating dynamic objects (like vehicles) from static elements (buildings, roads), using multiple neural fields for enhanced re…

New scaling laws for efficient floating-point quantization training in LLMs are presented, showing optimal bit allocation and critical data size.

VideoMaker achieves high-fidelity zero-shot customized video generation by cleverly harnessing the inherent power of video diffusion models, eliminating the need for extra feature extraction and injec…

DRT-01 leverages long chain-of-thought reasoning to significantly boost machine translation quality, particularly for complex sentences with metaphors and similes, achieving substantial improvements o…

This study reveals that gist token-based context compression in LLMs, while effective for some tasks, suffers from key failure patterns. The authors propose fine-grained autoencoding and segment-wise…

FreeSplatter: a novel feed-forward framework reconstructs high-quality 3D scenes from uncalibrated sparse-view images, estimating camera poses in seconds.

EMOv2 achieves state-of-the-art performance in various vision tasks using a novel Meta Mobile Block, pushing the 5M parameter lightweight model frontier.

Divot: A novel diffusion-powered video tokenizer enables unified video comprehension & generation with LLMs, surpassing existing methods.

NVComposer: A novel generative NVS model boosts synthesis quality by implicitly inferring spatial relationships from multiple sparse, unposed images, eliminating reliance on external alignment.

Boosting LLMs’ reasoning: A novel token-level contrastive estimation method automatically identifies and penalizes critical tokens leading to errors, significantly enhancing reasoning accuracy.

Self-VerIfication length Policy (SVIP) dynamically adjusts speculative decoding draft lengths based on token difficulty, achieving up to 20% faster large language model inference.

Low-bit quantization excels for undertrained LLMs but struggles with fully-trained ones; new scaling laws reveal this, directing future research.

AnchorCrafter animates cyber-anchors selling products via human-object interacting video generation, achieving high visual fidelity and controllable interactions.

Morph: a novel motion-free physics optimization framework drastically enhances human motion generation’s physical plausibility using synthetic data, achieving state-of-the-art quality.

Insight-V: A multi-agent system enhances multi-modal LLMs’ visual reasoning by generating high-quality long-chain reasoning data and employing a two-stage training pipeline, achieving significant perf…

StdGEN: Generate high-quality, semantically decomposed 3D characters from a single image in minutes, enabling flexible customization for various applications.

Tencent unveils Hunyuan-Large, a groundbreaking open-source MoE LLM boasting 389B parameters and 52B activated parameters, surpassing existing models in performance across various benchmarks.