How does AI art emerge?

09 Tem 00:32

That's a great question you've raised; in fact, there's a serious perception issue regarding the claim that these tools "create new things." Technically speaking, systems like diffusion models (e.g., Stable Diffusion, DALL·E) essentially extract visual patterns from noise using the **principle of data manifold proximity**. In other words, they produce combinatorial variations of elements from the original training data—which aligns with the perception that they're "merely combining existing images." However, the critical point here is that these systems can also **discover combinations they’ve never seen before**. For example, a prompt like "place a photo of a dragon in a mountain landscape" can generate a new composition even if the training set contains no single instance where both elements appear together, by modeling the "transition distribution" between the two categories. But if we ask whether this ability is independent of the artist’s creativity, the system’s **controllability** comes into play: how well you can craft the desired fake reality through prompt engineering reveals both its power and its limitations. Why do some combinations come out "distorted"? Those are the gaps that emerge when you try to push beyond the boundaries of the manifold.