news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

①through machine learning, researchers have enabled the gamengen model to successfully generate the 90s first-person shooter game "doom" in real time; ②although the finished product is still rough, the research further demonstrates the prospects for ai-generated games in the future; ③nvidia ceo jensen huang once predicted that we will be able to see games generated entirely by ai in about 5-10 years.

cailianshe news, august 29 (editor: shi zhengcheng)the era of real-time game generation by ai models has quietly arrived upon us.

this week, researchers from google and tel aviv university published a paper titleddiffusion model is a real-time game engine》, which introduces gamengen, the first game engine in computer history that is fully supported by a neural network model.

(source: github)

the researchers wrote in their paper:today, video games are programmed by humans, and gamengen is a partial proof-of-concept for a new paradigm for game engines — games will become weights of neural models rather than lines of code.。”

to put it in a more understandable way, all current video games are pre-designed, and developers need to write code, prepare game text and texture models, and then place them on the game map-the rendering and status update of the game screen depends on manually edited rules. but the gamengen model opens up a completely different idea:use ai to generate models and calculate and generate game screens in real time based on the player's actions and reactions。

in the demonstration, the researchers used machine learning to enable the gamengen model to successfully generate the 90s first-person shooter game "doom" in real time. the video shows that in the ai-generated game, players can turn and fire weapons in the scene, and can accurately reflect the number of remaining bullets, the remaining health after being attacked, and whether the conditions required to open the next level are met.

(source: demo video)

it should be noted that the series of images seen above are completely generated by ai in real time.the latest progress also shows that after successfully generating text, images, audio and short videos, ai models may have the ability to generate game scenes, which requires significantly higher logic, coherence and real-time interaction.

how did they do it?

the research team said that in order to train this ai that can generate games in real time, it is first necessary to train a reinforcement learning (rl) agent to play the game, and then use the recorded clips to train the generated diffusion model to predict the next picture based on past pictures and player actions. this is why the ai-generated game can show changes in health and ammunition, as well as animations of enemies being attacked.

the greater challenge is to keep the ai-generated images temporally and logically coherent. to mitigate autoregressive drift during inference, the researchers corrupted the context frames by adding gaussian noise to the encoded frames during training, allowing the ai ​​to correct the information sampled in the previous frames, thereby maintaining the stability of image generation over a long period of time.

(source: research paper)

the researchers revealed thatrunning this model only requires a single tpu (google's self-developed ai processor) to achieve a generation speed of 20 frames per second.。

of course, the above paragraphs also show the limitations of gamengen: this ai must rely on inputting existing games (or text, pictures and other materials) to generate games.

dr. jim fan, senior research manager & head of embodied intelligence group at nvidia, commented on social media:gamengen is more like a neural radiance field (nerf) than a video generation model. nerf generates a 3d representation of a scene by taking images of the scene from different angles. but this also means that the model does not have the ability to generalize and cannot "imagine" new scenes. this is also the difference between gamengen and sora: it cannot generate new scenes or interaction mechanisms.

(source: x)

the researchers also mentioned this in the paper, explaining that with this technology,in the future, game developers will be able to create new games through "text descriptions and sample images", and it will be possible for people to transform a set of refined images into new playable levels or characters for an existing game based solely on examples rather than programming skills.

huang renxun: games completely generated by ai will appear in 5-10 years

real-time gaming rendered by ai is not a new idea. when the latest generation of blackwell architecture chips was released in march this year, nvidia ceo jensen huang predicted thatin about 5-10 years, we will see games generated entirely by ai.。

in fact, it’s not just the google team that’s moving in this direction.when openai first released the sora demo this year, it also demonstrated its ability to simulate the pixel game "minecraft"。

(source: openai)

the latest development also coincides with the "persuasion to quit" remarks made by cai haoyu, the former chairman of mihoyo, which have recently sparked heated discussions.

cai haoyu publicly stated this week thataigc has revolutionized game development, now it just takes time for this phenomenon to fully unfoldhe believes that in the future, only two types of game developers will continue to work in the industry: the top 0.0001% of geniuses, and the 99% of business enthusiasts who create games to meet their own needs. as for the remaining "ordinary to professional" game developers, it is better for everyone to change careers as soon as possible.

(source: internet)