This article reviews notable AI research papers published in Week 48 of 2024 (24W48), covering image generation, 3D/4D content creation, language models, and evaluation.

Image Generation: Style-Friendly SNR Sampler improves style-driven generation by shifting the signal-to-noise ratio distribution toward higher noise levels during fine-tuning, enabling effective learning of personal artistic styles including watercolor, flat cartoon, 3D rendering, and meme aesthetics. OminiControl integrates image conditioning into pretrained Diffusion Transformer models with only ~0.1% additional parameters, unifying subject-driven generation and spatially-aligned conditioning tasks.

3D/4D Content: Material Anything provides automated end-to-end physically-based material generation for 3D objects using a triple-head architecture with rendering loss, handling diverse lighting conditions without case-specific optimization. CAT4D generates 4D scenes from monocular video using a multi-view video diffusion model trained on diverse dataset combinations.

Language Models: TÜLU 3, a fully open model, surpasses GPT-4o-mini and Claude 3.5-Haiku on instruction-following benchmarks through improved RLHF data curation and training recipes. Star Attention achieves efficient long-sequence processing through block-sparse attention patterns with global anchor tokens, enabling near-linear scaling for context lengths up to 1M tokens. ShowUI specializes in GUI screenshot understanding for task automation. ROICtrl enables precise instance-level attribute control in image generation. LLM-as-a-judge research provides insights into calibration, consistency, and position bias of model-based evaluation.