# Cracking the Code: How Diffusion Models are Speeding Up AI > Source: > Published: 2026-07-14 15:23:59+00:00 # Cracking the Code: How Diffusion Models are Speeding Up AI Diffusion generative models are getting faster by tapping into hidden low-dimensional structures. DDIM and DDPM lead the charge with groundbreaking convergence rates. JUST IN: Diffusion generative models are making waves by speeding up [sampling](/glossary/sampling) times through some wild low-dimensional wizardry. The denoising diffusion implicit model (DDIM) and its cousin, the denoising diffusion probabilistic model (DDPM), are at the forefront of this evolution. ## What's the Deal? Let's break it down. These models are achieving iteration complexities that are off the charts. We're talking about performance pegged at the order of k/ε (with a cheeky log factor on the side), where ε is your total variation distance precision, and k is a mysterious intrinsic dimension of the target distribution. It's a mouthful, but basically, these models are getting faster at handling complex data. The kicker? They're doing it even when the score functions, basically the model’s compass, are learned from data, not just handed on a silver platter. That means they're not just fast, they're adaptive. They're rolling with the punches and still coming out ahead. ## Why Should You Care? Sources confirm: this isn't just theoretical babble. The labs are scrambling to catch up. These models could change AI by enabling faster and more efficient sampling, which means more power and speed in generating complex data. And just like that, the leaderboard shifts, with DDIM and DDPM setting new standards. Here's the real question: if these models can adapt and perform under less-than-ideal conditions, what does that mean for future AI development? It means we're looking at a new era where AI can learn and adapt more freely. The possibilities are endless, and the competition will be fierce. ## The Takeaway Kernel-based score estimators are the unsung heroes here, providing the finite-sample guarantees that make all this possible. It's like they've cracked the code on using low-dimensional structure to their advantage. This research doesn't just push the boundaries, it obliterates them. Forget playing it safe, it's time to rethink how we approach AI models. In a world where speed and adaptability are king, DDIM and DDPM have just set the bar higher. The future of AI isn't just faster, it's smarter. Get AI news in your inbox Daily digest of what matters in AI.