User contributions for Felipefelixarias

00:2100:21, 9 September 2024 diff hist +995‎ N Translations:Diffusion Models Are Real-Time Game Engines/28/zh ‎ Created page with "我们重新利用预训练的文本到图像扩散模型 Stable Diffusion v1.4（Rombach 等人，[https://arxiv.org/html/2408.14837v1#bib.bib26 2022]）。我们将模型 <math>f_{\theta}</math> 置于轨迹 <math>T \sim \mathcal{T}_{agent}</math> 的条件下，即在之前的动作 <math>a_{< n}</math> 和观察（帧） <math>o_{< n}</math> 的序列条件下，并移除所有文本条件。具体来说，为了以动作为条件，我们仅需学习将每个动作..." current
00:2100:21, 9 September 2024 diff hist −392‎ Diffusion Models Are Real-Time Game Engines/zh ‎ Created page with "现在，我们训练一个生成扩散模型，该模型以在前一阶段收集的代理轨迹<math>\mathcal{T}_{agent}</math>（行动和观察）作为条件。"
00:2100:21, 9 September 2024 diff hist +164‎ N Translations:Diffusion Models Are Real-Time Game Engines/27/zh ‎ Created page with "现在，我们训练一个生成扩散模型，该模型以在前一阶段收集的代理轨迹<math>\mathcal{T}_{agent}</math>（行动和观察）作为条件。" current
00:2000:20, 9 September 2024 diff hist +36‎ N Translations:Diffusion Models Are Real-Time Game Engines/26/zh ‎ Created page with "=== 3.2 训练生成扩散模型 ===" current
00:2000:20, 9 September 2024 diff hist +281‎ N Translations:Diffusion Models Are Real-Time Game Engines/25/zh ‎ Created page with "我们在整个训练过程中记录了代理的训练轨迹，其中涵盖了不同技能水平的游戏。这组记录的轨迹构成了我们的<math>\mathcal{T}_{agent}</math>数据集，用于训练生成模型（见第[https://arxiv.org/html/2408.14837v1#S3.SS2 3.2]节）。" current
00:2000:20, 9 September 2024 diff hist +748‎ N Translations:Diffusion Models Are Real-Time Game Engines/24/zh ‎ Created page with "我们的最终目标是让人类玩家与我们的仿真进行互动。为此，第[https://arxiv.org/html/2408.14837v1#S2 2]节中的策略<math>\pi</math>即为“人类游戏策略”。由于我们无法直接大规模地从中取样，因此我们首先通过教一个自动代理来玩游戏，以此来近似人类游戏。与典型的强化学习设置不同，该设置旨在最大化游戏得分，我们的目标是生成与人类游戏类似的训练数据，或..." current
00:2000:20, 9 September 2024 diff hist +42‎ N Translations:Diffusion Models Are Real-Time Game Engines/23/zh ‎ Created page with "=== 3.1 通过代理进行数据收集 ===" current
00:2000:20, 9 September 2024 diff hist +141‎ N Translations:Diffusion Models Are Real-Time Game Engines/22/zh ‎ Created page with "center|thumb|900x900px|图3：GameNGen方法概览。为了简洁起见，省略了v预测的详细信息。" current
00:2000:20, 9 September 2024 diff hist −135‎ Diffusion Models Are Real-Time Game Engines/zh ‎ Created page with "GameNGen（发音为“游戏引擎”）是一个生成扩散模型，它能够在第[https://arxiv.org/html/2408.14837v1#S2 2]节的设置下学习模拟游戏。为了收集该模型的训练数据，我们首先使用教师强制目标训练一个独立的模型与环境进行交互。这两个模型（代理和生成模型）依次进行训练。在训练过程中，代理的全部行为和观察语料 <math>\mathcal{T}_{agent}</math> 被保留下来，并在第二..."
00:2000:20, 9 September 2024 diff hist +574‎ N Translations:Diffusion Models Are Real-Time Game Engines/21/zh ‎ Created page with "GameNGen（发音为“游戏引擎”）是一个生成扩散模型，它能够在第[https://arxiv.org/html/2408.14837v1#S2 2]节的设置下学习模拟游戏。为了收集该模型的训练数据，我们首先使用教师强制目标训练一个独立的模型与环境进行交互。这两个模型（代理和生成模型）依次进行训练。在训练过程中，代理的全部行为和观察语料 <math>\mathcal{T}_{agent}</math> 被保留下来，并在第二..." current
00:2000:20, 9 September 2024 diff hist −56‎ Diffusion Models Are Real-Time Game Engines/zh ‎ Created page with "== 2 互动世界仿真 =="
00:2000:20, 9 September 2024 diff hist +16‎ N Translations:Diffusion Models Are Real-Time Game Engines/20/zh ‎ Created page with "== 3 GameNGen ==" current
00:2000:20, 9 September 2024 diff hist −574‎ Diffusion Models Are Real-Time Game Engines/zh ‎ Created page with "我们总是使用教师强迫目标来训练我们的生成模型。给定一个模拟分布函数 <math>q</math>，可以通过自回归地采样观测值来模拟环境 <math>\mathcal{E}</math>。"
00:1900:19, 9 September 2024 diff hist +199‎ N Translations:Diffusion Models Are Real-Time Game Engines/19/zh ‎ Created page with "我们总是使用教师强迫目标来训练我们的生成模型。给定一个模拟分布函数 <math>q</math>，可以通过自回归地采样观测值来模拟环境 <math>\mathcal{E}</math>。" current
00:1900:19, 9 September 2024 diff hist +1,196‎ N Translations:Diffusion Models Are Real-Time Game Engines/18/zh ‎ Created page with "给定输入交互环境 <math>\mathcal{E}</math> 和初始状态 <math>s_{0} \in \mathcal{S}</math>，一个“交互世界模拟”是一个“模拟分布函数” <math>q \left( o_{n} \,|\, \{o_{< n}, a_{\leq n}\} \right), \; o_{i} \in \mathcal{O}, \; a_{i} \in \mathcal{A}</math>。给定观测值之间的距离度量 <math>D: \mathcal{O} \times \mathcal{O} \rightarrow \mathbb{R}</math>，一个“策略”，即给定过去动作和观测的代理动作分布 <math>..." current
00:1900:19, 9 September 2024 diff hist +340‎ N Translations:Diffusion Models Are Real-Time Game Engines/17/zh ‎ Created page with "例如，在游戏 DOOM 中，<math>\mathcal{S}</math> 是程序的动态内存内容，<math>\mathcal{O}</math> 是渲染的屏幕像素，<math>V</math> 是游戏的渲染逻辑，<math>\mathcal{A}</math> 是按键和鼠标移动的集合，而 <math>p</math> 是基于玩家输入的程序逻辑（包括任何潜在的非确定性）。" current
00:1900:19, 9 September 2024 diff hist +436‎ N Translations:Diffusion Models Are Real-Time Game Engines/16/zh ‎ Created page with "一个''交互环境''<math>\mathcal{E}</math>由一个潜在状态空间<math>\mathcal{S}</math>、一个潜在空间的部分投影空间<math>\mathcal{O}</math>、一个部分投影函数<math>V: \mathcal{S} \rightarrow \mathcal{O}</math>、一组动作<math>\mathcal{A}</math>，以及一个转移概率函数<math>p \left( s^{\prime} \,|\, a, s \right)</math>，使得<math>s, s^{\prime} \in \mathcal{S}, a\in \mathcal{A}</math>。" current
00:1900:19, 9 September 2024 diff hist −66‎ Diffusion Models Are Real-Time Game Engines/zh ‎ Created page with "center|thumb|800x800px|图 2：GameNGen 与之前最先进的 DOOM 仿真的比较"
00:1900:19, 9 September 2024 diff hist +26‎ N Translations:Diffusion Models Are Real-Time Game Engines/15/zh ‎ Created page with "== 2 互动世界仿真 ==" current
00:1900:19, 9 September 2024 diff hist −229‎ Diffusion Models Are Real-Time Game Engines/zh ‎ Created page with "== 1 介绍 =="
00:1900:19, 9 September 2024 diff hist +115‎ N Translations:Diffusion Models Are Real-Time Game Engines/14/zh ‎ Created page with "center|thumb|800x800px|图 2：GameNGen 与之前最先进的 DOOM 仿真的比较" current
00:1900:19, 9 September 2024 diff hist +393‎ N Translations:Diffusion Models Are Real-Time Game Engines/13/zh ‎ Created page with "GameNGen 回答了在通往游戏引擎新范式的道路上一个重要的问题，即游戏可以自动生成，就像近年来神经模型生成图像和视频一样。仍然存在关键问题，例如如何训练这些神经游戏引擎，以及如何有效地创建游戏，包括如何最佳地利用人类输入。尽管如此，我们对这种新范式的可能性感到非常兴奋。" current
00:1900:19, 9 September 2024 diff hist −247‎ Diffusion Models Are Real-Time Game Engines/zh ‎ Created page with "在这项工作中，我们证明答案是肯定的。具体来说，我们展示了一款复杂的视频游戏——标志性游戏《DOOM》，可以在神经网络（开放式 Stable Diffusion v1.4 的增强版（Rombach 等人，[https://arxiv.org/html/2408.14837v1#bib.bib26 2022]））上实时运行，同时获得与原始游戏相当的视觉质量。尽管这不是精确仿真，该神经模型能够执行复杂的游戏状态更新，例如统计生命值和弹..."
00:1900:19, 9 September 2024 diff hist +567‎ N Translations:Diffusion Models Are Real-Time Game Engines/12/zh ‎ Created page with "在这项工作中，我们证明答案是肯定的。具体来说，我们展示了一款复杂的视频游戏——标志性游戏《DOOM》，可以在神经网络（开放式 Stable Diffusion v1.4 的增强版（Rombach 等人，[https://arxiv.org/html/2408.14837v1#bib.bib26 2022]））上实时运行，同时获得与原始游戏相当的视觉质量。尽管这不是精确仿真，该神经模型能够执行复杂的游戏状态更新，例如统计生命值和弹..." current
00:1800:18, 9 September 2024 diff hist +81‎ N Translations:Diffusion Models Are Real-Time Game Engines/11/zh ‎ Created page with "一个实时运行的神经模型是否能够以高质量模拟复杂的游戏？" current
00:1800:18, 9 September 2024 diff hist +568‎ N Translations:Diffusion Models Are Real-Time Game Engines/10/zh ‎ Created page with "有几项重要研究（Ha & Schmidhuber，[https://arxiv.org/html/2408.14837v1#bib.bib10 2018]；Kim 等人，[https://arxiv.org/html/2408.14837v1#bib.bib16 2020]；Bruce 等人，[https://arxiv.org/html/2408.14837v1#bib.bib7 2024]）（见第[https://arxiv.org/html/2408.14837v1#S6 6]节）使用神经模型来模拟交互式视频游戏。然而，这些方法大多在模拟游戏的复杂性、仿真速度、长时间的稳定性或视觉质量等方面存在局限性..." current
00:1800:18, 9 September 2024 diff hist −213‎ Diffusion Models Are Real-Time Game Engines/zh ‎ Created page with "近年来，生成模型在根据文本或图像等多模态输入生成图像和视频方面取得了重大进展。在这一浪潮的前沿，扩散模型成为非语言媒体生成的事实标准，如 Dall-E（Ramesh 等人，[https://arxiv.org/html/2408.14837v1#bib.bib25 2022]）、Stable Diffusion（Rombach 等人，[https://arxiv.org/html/2408.14837v1#bib.bib26 2022]）和 Sora（Brooks 等人，[https://arxiv.org/html/2408.14837v1#bib.bib6 2024]）。乍一看，..."
00:1800:18, 9 September 2024 diff hist +924‎ N Translations:Diffusion Models Are Real-Time Game Engines/9/zh ‎ Created page with "近年来，生成模型在根据文本或图像等多模态输入生成图像和视频方面取得了重大进展。在这一浪潮的前沿，扩散模型成为非语言媒体生成的事实标准，如 Dall-E（Ramesh 等人，[https://arxiv.org/html/2408.14837v1#bib.bib25 2022]）、Stable Diffusion（Rombach 等人，[https://arxiv.org/html/2408.14837v1#bib.bib26 2022]）和 Sora（Brooks 等人，[https://arxiv.org/html/2408.14837v1#bib.bib6 2024]）。乍一看，..." current
00:1800:18, 9 September 2024 diff hist −259‎ Diffusion Models Are Real-Time Game Engines/zh ‎ Created page with "计算机游戏是围绕以下“游戏循环”手动制作的软件系统：(1) 收集用户输入，(2) 更新游戏状态，(3) 将其渲染为屏幕像素。这个游戏循环以很高的帧率运行，为玩家营造出一个交互式虚拟世界的假象。这种游戏循环通常在标准计算机上运行，尽管也有许多在定制硬件上运行游戏的惊人尝试（例如，标志性游戏《毁灭战士》曾在烤面包机、微波炉、跑步机、照..."
00:1800:18, 9 September 2024 diff hist +848‎ N Translations:Diffusion Models Are Real-Time Game Engines/8/zh ‎ Created page with "计算机游戏是围绕以下“游戏循环”手动制作的软件系统：(1) 收集用户输入，(2) 更新游戏状态，(3) 将其渲染为屏幕像素。这个游戏循环以很高的帧率运行，为玩家营造出一个交互式虚拟世界的假象。这种游戏循环通常在标准计算机上运行，尽管也有许多在定制硬件上运行游戏的惊人尝试（例如，标志性游戏《毁灭战士》曾在烤面包机、微波炉、跑步机、照..."
00:1800:18, 9 September 2024 diff hist +14‎ N Translations:Diffusion Models Are Real-Time Game Engines/7/zh ‎ Created page with "== 1 介绍 ==" current
00:1800:18, 9 September 2024 diff hist +224‎ N Translations:Diffusion Models Are Real-Time Game Engines/6/zh ‎ Created page with "center|thumb|800x800px|图 1：一名玩家正在 '''GameNGen''' 上以 20 FPS 的速度游玩 DOOM。请参见 [https://gamengen.github.io/ https://gamengen.github.io] 获取演示视频。" current
00:1800:18, 9 September 2024 diff hist −200‎ Diffusion Models Are Real-Time Game Engines/zh ‎ Created page with "我们介绍了''GameNGen''，这是第一个完全由神经模型驱动的游戏引擎，能够在长轨迹上与复杂环境进行高质量的实时交互。GameNGen 可以在单个 TPU 上以每秒超过 20 帧的速度交互模拟经典游戏 DOOM。下一帧预测的 PSNR 为 29.4，与有损 JPEG 压缩相当。在区分游戏短片和模拟片段方面，人类评分员的表现仅略好于随机概率。GameNGen 的训练分为两个阶段：(1) 一个强化学..."
00:1800:18, 9 September 2024 diff hist +684‎ N Translations:Diffusion Models Are Real-Time Game Engines/5/zh ‎ Created page with "我们介绍了''GameNGen''，这是第一个完全由神经模型驱动的游戏引擎，能够在长轨迹上与复杂环境进行高质量的实时交互。GameNGen 可以在单个 TPU 上以每秒超过 20 帧的速度交互模拟经典游戏 DOOM。下一帧预测的 PSNR 为 29.4，与有损 JPEG 压缩相当。在区分游戏短片和模拟片段方面，人类评分员的表现仅略好于随机概率。GameNGen 的训练分为两个阶段：(1) 一个强化学..." current
00:1800:18, 9 September 2024 diff hist −58‎ Diffusion Models Are Real-Time Game Engines/zh ‎ Created page with "====== 摘要 ======"
00:1800:18, 9 September 2024 diff hist +20‎ N Translations:Diffusion Models Are Real-Time Game Engines/4/zh ‎ Created page with "====== 摘要 ======" current
00:0400:04, 9 September 2024 diff hist +62,004‎ N Diffusion Models Are Real-Time Game Engines/zh ‎ Created page with "'''项目网站：''' [https://gamengen.github.io/ https://gamengen.github.io]"
00:0400:04, 9 September 2024 diff hist +78‎ N Translations:Diffusion Models Are Real-Time Game Engines/3/zh ‎ Created page with "'''项目网站：''' [https://gamengen.github.io/ https://gamengen.github.io]" current
00:0300:03, 9 September 2024 diff hist +88‎ N Translations:Diffusion Models Are Real-Time Game Engines/2/zh ‎ Created page with "'''ArXiv链接：''' [https://arxiv.org/abs/2408.14837 https://arxiv.org/abs/2408.14837]" current
00:0300:03, 9 September 2024 diff hist +158‎ N Translations:Diffusion Models Are Real-Time Game Engines/1/zh ‎ Created page with "'''作者：''' Dani Valevski（谷歌研究）、Yaniv Leviathan（谷歌研究）、Moab Arar（特拉维夫大学）、Shlomi Fruchter（谷歌 DeepMind）" current

18:2618:26, 7 September 2024 diff hist −11‎ Translations:Welcome/6/zh ‎No edit summary
18:2318:23, 7 September 2024 diff hist −12‎ m Welcome ‎ Changed headshot to frameless Tag: Visual edit: Switched
07:0707:07, 7 September 2024 diff hist −1‎ Translations:Welcome/6/zh ‎No edit summary
06:5406:54, 7 September 2024 diff hist +119‎ Welcome/zh ‎ Created page with "right|frameless '''<big>Marovi 是一个面向人工智能研究和资源的多语言平台。我们翻译最重要的人工智能研究论文，同时汇集了公开可用的最大多语言人工智能相关术语词汇表。</big>'''"
06:5306:53, 7 September 2024 diff hist +79‎ Translations:Welcome/6/zh ‎No edit summary
06:5106:51, 7 September 2024 diff hist 0‎ Welcome ‎No edit summary Tag: Visual edit: Switched
06:4806:48, 7 September 2024 diff hist +79‎ Welcome ‎ Added Felipe Headshot Tag: Visual edit
06:4406:44, 7 September 2024 diff hist +41‎ N File:FelipeFelixArias2024.jpg ‎No edit summary current
06:3506:35, 7 September 2024 diff hist −38‎ Diffusion Models Are Real-Time Game Engines/es ‎ Created page with "Evaluamos el impacto de cambiar el número <math>N</math> de observaciones pasadas en el contexto de condicionamiento entrenando modelos con <math>N \in \{ 1,2,4,8,16,32,64\}</math> (recordemos que nuestro método utiliza <math>N = 64</math>). Esto afecta tanto al número de fotogramas históricos como al de acciones. Entrenamos los modelos durante 200.000 pasos manteniendo congelado el decodificador y evaluamos en trayectorias de prueba de 5 niveles. Véanse los resulta..." current
06:3406:34, 7 September 2024 diff hist +359‎ N Translations:Diffusion Models Are Real-Time Game Engines/45/es ‎ Created page with "Observamos que es sencillo aumentar aún más la tasa de generación de imágenes de manera sustancial al paralelizar la generación de varios fotogramas en hardware adicional, similar a la técnica clásica de Nvidia SLI Alternate Frame Rendering (AFR). Al igual que con AFR, la tasa real de simulación no aumentaría ni se reduciría el retardo de entrada." current

9 September 2024

7 September 2024