🏢 KAIST

ORIGEN: First zero-shot 3D orientation grounding in text-to-image generation.

Fixing fine-tuned diffusion models! By using richer, unconditional priors, they generate better images and videos.

Inference-time scaling for flow models enhances alignment with user preferences via stochastic generation and budget allocation.

New ‘Silent Branding Attack’ poisons text-to-image models, embedding brand logos without text prompts, raising ethical issues for image generation tools.

SynCoS: Synchronized sampling generates high-quality & coherent long videos from text, without extra training!

Sketch-of-Thought(SoT) reduces LLM token usage by up to 76% while maintaining (or improving) accuracy via cognitive-inspired sketching.

LLMs fail to act safely when considering user-specific safety standards, which were made to be solved via new benchmark.

SafeRoute efficiently enhances LLM safety by adaptively using smaller and larger safety guard models, maximizing accuracy while minimizing costs.

Edit many objects at once in videos! MIVE does it accurately without affecting other areas, a big step for AI video editing.

BootComp: generate realistic human images wearing multiple garments using a novel synthetic data pipeline & diffusion model, enabling diverse applications like virtual try-on.

CoordTok: a novel video tokenizer drastically reduces token count for long videos, enabling memory-efficient training of diffusion models for high-quality, long video generation.