---
schema: 1
type: term
title: '**next-frame prediction**'
description: >-
  **Next-frame prediction** is a model approach for video in which the system
  generates the next frame of a video from the frames that came before it. In
  this channel's material, it underpins real-time **video-to-video AI**:
  predicting frames fast enough (in tens of milliseconds) is what makes an
  explorable, interactive video-game world possible.
canonical: >-
  https://liquidgold.cc/channel/everyinc/explore/term/next-frame-prediction-89f04deafc7f
markdown: >-
  https://liquidgold.cc/channel/everyinc/explore/term/next-frame-prediction-89f04deafc7f.md
llms: null
parent: 'https://liquidgold.cc/channel/everyinc/llms.txt'
channel: everyinc
updated: '2026-10-23T22:12:35.346Z'
language: null
children: []
---


# \*\*next-frame prediction\*\*

Canonical: https://liquidgold.cc/channel/everyinc/explore/term/next-frame-prediction-89f04deafc7f

Parent: [Channel index](https://liquidgold.cc/channel/everyinc/llms.txt)

## Definition

**Next-frame prediction** is a model approach for video in which the system generates the next frame of a video from the frames that came before it. In this channel's material, it underpins real-time **video-to-video AI**: predicting frames fast enough (in tens of milliseconds) is what makes an explorable, interactive video-game world possible.

## How this channel uses the term\{' '\}In the conversation with Dean Leitersdorf, co-founder and CEO of Decart, **next-frame prediction** is the architectural idea behind real-time **video-to-video AI**: the model generates a video world by continuously producing the next frame from what came before. Getting that loop fast enough — frame generation measured in *tens of milliseconds* — is what turns video prediction into an explorable, interactive game world rather than a pre-rendered clip. The material frames this as a departure from the LLM paradigm: it is one of the newer architectures and capabilities explored in the topic *Beyond LLMs: New AI Architectures and Capabilities*, alongside **energy-based models** and **hybrid diffusion–autoregressive video models**.## Architecture behind real-time video\{' '\}## Why it matters\{' '\}\{'  '\}- **Next-frame prediction** = building a video world one predicted frame at a time.\{'  '\}- Speed is the crux: milliseconds-per-frame is what enables *real-time* video-to-video AI.\{'  '\}- The channel presents it as an example of architectures designed for problems beyond what LLMs address — here, generating explorable, interactive worlds.