Why AI images go soft, and a new preprint that blames the training loop

Every AI image generator has a tell: a faint softness around the edges, as if the picture forgot to commit. A new preprint argues that softness is not an accident of scale. It is a property of how the model is trained.

The paper, "Explorative Modeling" (arXiv 2607.27372) and matching author blog post, proposes a third axis for pretraining, beyond parameters and data, built around a "best of K" rule. At each training step the model generates K candidate outputs, scores them against the target, and updates on the lowest-loss match. The author calls the family of resulting models "Explorative Models" (XMs); a project site and open-source code release accompany the preprint.

The intuition is closer to darts than to deep learning. When several answers are all valid, the single best prediction is their average. That average looks like nothing in particular. A model trained against an averaged target is, in effect, being told to imitate a thing that does not exist. The author's framing: blur is the visible signature of that mismatch.

In headline numbers, the preprint reports 6.2× sample efficiency and 4.1× FLOP efficiency, meaning the same score at roughly one-sixth the data and one-fourth the compute, and says the gains grow rather than shrink with scale: 7%→36% as data rises, 13%→23% as parameters rise. On end-to-end generation, XMs are reported to match diffusion on control tasks with up to 256× less inference compute, a category of inputs (edges, edges-to-image, segmentation maps) where the standard diffusion baseline is unusually strong.

Best-of-K is not a foreign idea. RLHF pipelines already sample several candidates and train on the best, and inference-time best-of-N is a common trick for sharpening model outputs. The novelty is doing the selection at the training step, changing the loss itself rather than the policy that runs on top of it. The author positions the work as a training-objective critique that applies across images, video, and language, not a tweak to one pipeline.

It is not the only reading. A top comment on the Hacker News thread argues the "blur problem" is a misunderstanding of how standard generative models already work. They predict distributions rather than point estimates, the commenter writes; what XMs add is best-of-K, closer to a hard version of a latent-variable model than a fix to a foundational flaw. The critique does not dispute the efficiency numbers. It disputes the diagnosis, and it is anonymous, which limits how much weight a reader should give it.

The source itself narrows the claim. The preprint is single-author, not peer-reviewed, and lists no co-authors, lab, or independent reproduction. The efficiency multipliers are the author's, not the field's. The open code release lowers the bar for an outside check, and an active Hacker News thread shows practitioners already poking at the framing.

That is the part worth sitting with. If blur is indeed a property of the training objective rather than a residue of scale, the next wave of generative-AI efficiency gains may come from changing what the model is imitating, not how big the model is. The dartboard is small. The question it raises is not.