You can play Genshin Impact just by moving your mouth! Use AI to switch characters and attack enemies. Netizen: 'Ayaka, use Kamiri-ryu Frost Destruction'

Speaking of domestic games that have become popular all over the world in the past two years, Genshin Impact definitely takes the cake.

According to this year’s Q1 quarter mobile game revenue survey report released in May, “Genshin Impact” firmly won the first place among card-drawing mobile games with an absolute advantage of 567 million U.S. dollars. This also announced that “Genshin Impact” 》In just 18 months after its launch, the total revenue on the mobile platform alone exceeded US$3 billion (approximately RM13 billion).

Now, the last 2.8 island version before the opening of Xumi is long overdue. After a long draft period, there are finally new plots and areas to play.

But I don’t know how many “Liver Emperors” there are. Now that the island has been fully explored, grass has begun to grow again.

There are 182 treasure chests in total and 1 Mora box (not included)

There is nothing to be afraid of during the long grass period, there is never a shortage in the Genshin Impact area Whole job.

No, during the long grass period, some players used XVLM wenet STARK to make a voice control project to play Genshin Impact.

For example, when he said "Use Tactic 3 to attack the fire slime in the middle", Zhongli first used a shield, Ling Hua did a step step and then said "Sorry", and the group destroyed 4 of them. Fire Slime.

Similarly, after saying "attack the Daqiuqiu people in the middle", Diona used E to set up a shield, Ling Hua followed up with an E and then 3A to clean up beautifully. Two large Qiuqiu people were lost.

As can be seen on the lower left, the entire process was done without any use of hands.

Digest Fungi said that he is an expert, he can save his hands when writing books in the future, and said that mom no longer has to worry about tenosynovitis from playing Genshin Impact!

The project is currently open source on GitHub:

GitHub link:

https://github.com/7eu7d7/genshin_voice_play

Good Genshin Impact, he was actually played as a Pokémon

Such a live-action project naturally attracted the attention of many Genshin Impact players.

For example, some players suggested that the design can be more neutral and directly use the character name plus the skill name. After all, the audience cannot know the instructions such as "Tactics 3" at the first time, and "Zhongli, Using "Centre of the Earth" makes it easy to substitute into the game experience.

Some netizens said that since they can give instructions to monsters, can they also give voice commands to characters, such as "Turtle, use Frost Destruction".

guiguidailyquestion.jpg

However, why do these instructions seem so familiar?

In response to this, the up owner "Schrödinger's Rainbow Cat" said that the speed of shouting skills may not be kept up, and the attack speed will be slower. This is why he A set is preset.

However, the output methods of some classic teams, such as "Wanda International" and "Lei Jiuwan Ban" are relatively fixed, and the preset attack sequence and mode seem to be It works.

Of course, in addition to making jokes, netizens are also brainstorming and putting forward many optimization suggestions.

For example, directly use "1Q" to let the character in position 1 magnify his moves, use "heavy" to express heavy attacks, and "dodge" to dodge. In this way, it will be easier and faster to issue instructions, and it may also be used to fight the abyss. .

Some expert players also said that this AI seems to "do not understand the environment very well", "the next step can be to consider adding SLAM", "to achieve 360-degree all-round target detection" ".

# The owner of up said that the next step is to "completely automatically refresh the base, teleport, defeat monsters, and receive rewards in one package". It seems that an automatic strengthening saint can also be added. The relic function will format the AI ​​if it is crooked.

The hard-core live-up master of Genshin Impact also published the "Tivat Fishing Guide"

As Digest Fungus said, Genshin Impact has never There is a lack of work, and this up owner "Schrödinger's Rainbow Cat" should be the most "hardcore" among them.

From "AI automatically places the maze" to "AI automatically plays", every mini-game produced by Genshin Impact can be said to be based on AI.

Among them, Wenzhijun also discovered the "AI automatic fishing" project (the good guy turns out to be you too). You only need to start the program, and all the fish in Teyvat can be bagged. thing.

Genshin Impact automatic fishing AI consists of two parts of the model: YOLOX and DQN:

YOLOX is used to locate and identify fish types and locate the landing point of the fishing rod;

DQN is used to adaptively control the click of the fishing process so that the intensity falls within the optimal area.

In addition, this project also uses transfer learning and semi-supervised learning for training. The model also contains some non-learnable parts that are implemented using traditional digital image processing methods such as opencv.

Project address:

https://github.com/7eu7d7/genshin_auto_fish

You still need to fish after the 3.0 update "Salted Fish Bow", I'll leave it to you!

Those "artifacts" that turn Genshin Impact into Pokémon

As a serious person, Digest Fungus feels it is necessary to educate everyone about the use of Genshin Impact voice project Several "artifacts".

X-VLM is a multi-granularity model based on the visual language model (VLM). It consists of an image encoder, a text encoder and a cross-modal encoder. The cross-modal encoder combines visual features and language Cross-modal attention between features to learn visual language alignment.

The key to learning multi-granularity alignment is to optimize X-VLM: 1) by combining bounding box regression loss and IoU loss to locate visual concepts in images given associated text; 2) at the same time, by contrast loss, matching Loss and masked language modeling losses for multi-granular alignment of text with visual concepts.

In fine-tuning and inference, X-VLM can leverage the learned multi-granularity alignment to perform downstream V L tasks without adding bounding box annotations in the input image.

Paper link:

https://arxiv.org/abs/2111.08276

WeNet is an end-to-end production-oriented Speech recognition toolkit, which introduces a unified two-pass (U2) framework and built-in runtime to handle streaming and non-streaming decoding modes in a single model.

Just at the beginning of July this year, WeNet launched version 2.0 and was updated in 4 aspects:

U2: Unified dual-channel framework with bidirectional attention decoder, including from Future context information of the right-to-left attention decoder to improve the representation ability of the shared encoder and the performance of the rescoring stage;

Introduces an n-gram-based language model and a WFST-based decoder to facilitate Understand the use of rich text data in production scenarios;

Designed a unified context bias framework that utilizes user-specific context to provide rapid adaptability for production and improve ASR accuracy in both "with LM" and "without LM" scenarios;

A unified IO is designed to support large-scale data for effective model training.

Judging from the results, WeNet 2.0 achieved a relative recognition performance improvement of up to 10% on various corpora compared with the original WeNet.

Paper link: https://arxiv.org/pdf/2203.15455.pdf

STARK is a spatio-temporal transformation network for visual tracking. Based on the baseline consisting of convolutional backbone, codec converter and bounding box prediction head, STARK has made 3 improvements:

Dynamic update template: use intermediate frames as dynamic templates to add to the input. Dynamic templates can capture appearance changes and provide additional time domain information;

score head: determine whether the dynamic template is currently updated;

Training strategy improvement: Divide training into two stages 1) In addition to score In addition to the head, use the baseline loss function to train. Ensure that all search images contain the target and allow the template to have positioning capabilities; 2) Use cross entropy to only optimize the score head and freeze other parameters at this time to allow the model to have positioning and classification capabilities.

Paper link:

https://openaccess.thecvf.com/content/ICCV2021/papers/Yan_Learning_Spatio-Temporal_Transformer_for_Visual_Tracking_ICCV_2021_paper.pdf

The above is the detailed content of You can play Genshin Impact just by moving your mouth! Use AI to switch characters and attack enemies. Netizen: 'Ayaka, use Kamiri-ryu Frost Destruction'. For more information, please follow other related articles on the PHP Chinese website!