Generative Model Architectures, Efficiency and Post-Training
Key Questions
What improvements does DAR bring to Diffusion Transformers?
DAR introduces dynamic routing for Diffusion Transformers, achieving +2.11 FID improvement and 8.75x speedup. It enhances efficiency in generative image modeling.
How does Lens compare to larger text-to-image models?
Lens is a 3.8B parameter text-to-image model that outperforms larger models while using only 19.3% of the compute. It demonstrates strong efficiency in generative architectures.
What is PiD and what speedups does it offer?
PiD enables 4-8x faster high-resolution decoding in under one second. It targets efficiency gains in post-training and inference for generative models.
Which methods address fine-grained video understanding in MLLMs?
SWIM improves fine-grained video understanding in multimodal large language models. It works alongside perception-reasoning decoupling techniques in VLMs and scaling theories treating LLMs as noisy channels.
What role does RankE play in discrete text-to-image models?
RankE uses co-evolution strategies to enhance performance in discrete text-to-image generation. It contributes to post-training efficiency alongside other methods like DAR and Lens.
DAR improves Diffusion Transformers (+2.11 FID, 8.75x speedup). Lens (3.8B T2I beats larger models at 19.3% compute), PiD (4-8x high-res decoding <1s), RankE (co-evolution for discrete T2I). SWIM for MLLM fine-grained video understanding; perception-reasoning decoupling in VLMs; LLMs as Noisy Channels scaling theory.