Generative Model

The output of a decision model is often a label or probability distribution over a finite action set, similar to a policy network in reinforcement learning, whereas generative models require autoregressive sampling over a large vocabulary.

Current large language models and generative models are fundamentally seeking statistical patterns in massive amounts of data

Generative models at high resolution often require additional upscaling or redrawing steps, and these steps may change the number of fingers, the position of accessories or clothing texture, so 4K cannot replace consistency checks, and instead amplifies inconsistency.

Using generative models as surrogates for simulators is becoming an active route in computational science.

The new method proposed in this research uses generative models to simultaneously fit the conditional distributions of two groups of data and draw synthetic samples, and the test statistic depends on both the observed data and the synthetic samples.

There is a "cat-and-mouse game" confrontation between AI-generated content detection models and generation models, and any single detection method is hard to keep effective over the long term.

For domains requiring physically credible and verifiable results (scientific research, engineering, safety-critical systems), generative models currently cannot replace genuine numerical simulation.

Generative models do not guarantee energy conservation, do not guarantee that fluids satisfy the Navier-Stokes equations, and only guarantee that the output is statistically close to the training data in visual terms.

Simulation pursues physical correctness; generative models pursue perceptual plausibility.

Traditional generative models struggle to guarantee that every output falls within a strictly defined feasible region and may generate paths that violate constraints.

The root cause of generative model defects lies in their frame-by-frame or token-by-token prediction of outputs, lacking explicit encoding of global physical constraints.

Visual simulation defects in current generative models include the absence of object permanence, distortion of fluid and soft-body dynamics, and long-sequence inconsistency.

When everyone uses the same set of models, homogenization of output content is almost inevitable

Generative models excel at producing content that appears reasonable and meets the expectations of most people, and their optimization objective naturally trends toward regression to the mean

Advanced generative models can produce millions of candidate protein sequences within hours, but experimental validation of each sequence may take weeks

In commercial advertising scenarios, the current best practice remains to overlay text as a separate layer during the post-production compositing stage, rather than relying on generative models to render it directly.

Current generative models are essentially limited to the association level—identifying statistical co-occurrence patterns in data—and lack causal reasoning capability.

生成模型的创造力和幻觉本质上是同一枚硬币的两面，彻底消除编造能力将使模型失去泛化和创造能力