🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Reverse KL typically suffers from the mode collapse problem. We propose minimizing the reverse KL in the diffusion space, which significantly improves mode coverage of the learned generative models.

🚀 We propose a new way to estimate the denoising covariance in diffusion models, which can lead to lower estimation error, better FID and likelihood tradeoff with fewer steps, and improved generation diversity. ArXiv link: arxiv.org/abs/2406.10808. More details below👇

Checkout this paper with some really interesting insights led by the excellent Jiajun He and Yuanqi Du - TLDR: Neural density samplers really need guidance imposed through Langevin annealing to make them work well

Checkout this work on training one step generative model, pre-training is key!