Scaling diffusion models at inference time

Anyone who generates AI images knows the drill: bumping your sampling steps from 20 to 50 sharpens the image, but cranking it to 100 or 200 usually yields zero improvement while burning through compute.

While LLMs unlocked dramatic gains through inference-time reasoning, diffusion models have stayed largely stuck behind this denoising bottleneck. A new paper published on arXiv explores how image models can finally scale their test-time compute beyond just stacking denoising steps.

Turning diffusion into a search game

Researchers observed that while diffusion models natively adjust compute via denoising steps, quality gains consistently plateau after just a few dozen steps.

Instead of overthinking the denoising process on a single path, the paper reframes diffusion sampling as a search problem aimed at discovering better starting noises.

The framework organizes this design space across two main pillars: verifiers that score intermediate outcomes and provide feedback, and search algorithms that navigate the space to find superior noise candidates.

Across both class-conditioned and text-conditioned benchmarks, the authors found that spending extra compute through this search mechanism substantially improved generated image quality. Because visual tasks vary widely, different verifier and algorithm pairings can be tailored to specific generation needs.

Why it matters

We think this is a clever and long-overdue transposition of the inference-time scaling playbook from language models to generative vision.

For creators and developers, this could eventually eliminate the exhausting manual routine of rerolling prompts dozens of times just to cherry-pick a decent result. Instead of humans acting as the external verifier, the model can spend extra seconds behind the scenes exploring noise candidates to deliver a refined output on the first try.

For the broader industry, it introduces an alternative battleground to raw model pre-training. If inference compute can reliably compensate for architecture limits, smaller and lighter diffusion models might achieve frontier-tier visual fidelity simply by thinking a bit longer before rendering.

The catch, of course, is inference cost and latency. Running verification loops and search algorithms on high-resolution image latents is computationally heavy, and commercial platforms will have to figure out whether users are actually willing to wait longer per image.