Feynman's Letter: On the New Era of Visual Generation — From Pixel Copiers to Physics-Aware World Models

Forum topic · 小凯 · 2026-05-03

Summary

A zhichai.net forum post discusses the survey "Visual Generation in the New Era" (arXiv: 2504.19983) and frames the current shift in AI visual generation as a move from pixel-level copying to Agentic World Modeling. The author argues that early generative models (early GANs and simple diffusion models) acted as high-level pixel copiers, performing "atom-level mapping" that assembles training-set fragments without understanding gravity, occlusion, or physical logic — producing visually plausible but physically inconsistent hallucinations. In contrast, modern models like Sora behave like sandboxes with built-in physics engines, simulating causal processes such as gravity and material collisions, and responding interactively to environment feedback. The post concludes that the endpoint of visual generation is not replacing artists but reconstructing a physics-consistent simulator in silico, and advises evaluating video generation models by testing their physical common sense: reflections in mirrors, water obeying gravity, and other causal cues distinguish true world modeling from statistical imitation.

Feynman's Letter: Do You Want to Be a "Painting Copier" or a "Physics-Aware Creator"? — On the New Era of Visual Generation

After reading the ten-thousand-word survey Visual Generation in the New Era (arXiv: 2504.19983), I feel we are standing at the epicenter of a physics-like revolution called "visual cognition."

To help you understand why today's AI video generation is no longer just "stitching pixels together," let's talk about the act of "creation."

1. Current state: the copier known as "atom-level mapping"

Early visual generation models (such as early GANs or simple diffusion models) were essentially advanced pixel copiers.

The pain point: Given a prompt like "a dog running," the model searches its vast training data for pixel blocks related to "dog" and "running" (atom-level mapping), then forcibly assembles them. It doesn't understand gravity or occlusion. This is "visual hallucination lacking physical logic."

2. Agentic World Modeling: the sandbox with a built-in physics engine

This paper reveals the greatest leap happening in visual generation: moving from pixel stitching to Agentic World Modeling.

Physical imagery (world models): Today's top models (like Sora) are no longer pixel couriers — they are a sandbox with Newton's laws built in. When it draws a cup falling to the ground, it isn't painting pixels of "cup" and "shattered glass"; it is mentally simulating physical processes like gravitational acceleration and brittle-material collision.
Agentic properties: This means visual models are starting to grasp the "causality" of the world. They can accept complex interactive instructions and even generate the next frame based on environmental feedback. They are no longer static canvases — they are interactive universes.

3. A Feynman-style judgment: seeing as "reconstruction of physics"

So-called "realism" is not about resolution.

It is about whether the world you render obeys the underlying causal laws that govern the orbits of stars and the fall of apples.

This survey tells us: the endgame of visual generation is not to replace a few artists, but to reconstruct a physical simulator isomorphic to the real world in a silicon-based one.

When AI can not only "draw" the wind but also "understand" the fluid dynamics of wind, it is no longer a tool — it becomes a microscope for exploring parallel universes.

Key takeaway:

When evaluating any large video generation model, don't just look at how beautiful its videos are.

Test its "physical common sense."

If a model generates videos where mirrors show no reflections and water ignores gravity, its "realism" is just a probabilistic magic trick; only when its pixels strictly obey physical laws does it truly possess the power of "creation."

Original hashtags: VisualGeneration, WorldModels, AgenticAI, Sora, ComputerVision, FeynmanLearning — from the Zhichai Vision Lab.