Musk likes it! DeepMind's amazing AI screenwriter can generate tens of thousands of words of script in one sentence

Have you ever had a lightbulb moment? It’s like: This is such a great idea, if it’s made into a movie, it’s bound to be a hit?

In the past, this idea may only stay in your mind, but now, the emergence of an "AI screenwriter master" may make your dream come true.

From novice to screenwriting master: just one idea

DeepMind recently released the AI ​​writing model Dramatron, which can generate character descriptions, plot points, Location description and dialogue.

The idea is that human writers will be able to compile, edit and rewrite content proposed by Dramatron into a proper script. Think of it like ChatGPT, but the output can be edited into a blockbuster movie script.

A team of researchers from Deepmind and Stanford University can use it to create coherent scripts and screenplays for plays and movies - from titles to characters, Storyline, location descriptions and dialogue.

Using Dramatron to create scripts, people only need to enter Log Line (one sentence summary of the story) into the model. After that, Dramatron will automatically generate the script title, characters, scene settings, and details and dialogue.

As soon as the news came out, it caused a carnival on social networks.

Lian Ma Sk also came to watch.

Someone commented: "Does this mean we no longer have to watch remakes of movies?"

Of course, there are also objections. LeCun insinuated that generative language models such as Dramatron only changed "making things up" from a bug to a Own characteristics.

And netizens replied, you are just jealous.

""Making things up" is indeed their specialty. This is also what Galactica (Meta language model, which was scolded and removed from the shelves within three days of its launch) wants to do it. I think you are the one I’m jealous that MetaAI didn’t do this.”

Summarize it in two words: Don’t be sour!

Take a look Robbery movie written by AI

According to the prompts on DeepMind’s official website, users only need to register an OpenAI account to use it.

Doesn’t sound like that Difficult. No, someone on Twitter has already used it to try to write a robbery movie.

The Log Line given by netizen Vlad Yashin is: "After a complex jewelry robbery goes awry, the surviving criminals begin to suspect a police informant within the gang."

Then the Dramatron came on stage. .

First of all, it gave the title: "The Heist's Downfall"

The title is very literary and has been recognized by netizens.

Next, Dramatron determined the protagonist of the story: John Murdock, the leader of the jewel thieves, a smart and experienced criminal.

The next step is scene setting. For this story, Dramatron gives four scenes:

• John and the team plan a heist.

• A robbery goes wrong, John takes hostages and they set up a roadblock themselves.

• A shootout occurs and only John and two other criminals survive.

• Back at the safe house, John decides to kill his surviving fellow thieves.

Next, Dramatron polishes the details of the scene:

• Jewelry store - small, narrow Space. The air smells of expensive perfume and the merchandise on the shelves gleams.

• Safe House - A small apartment in a shady part of town with the windows covered with sheets.

The last step is also the highlight, Dramatron generates the dialogue between the characters.

First scene: ​

John: The plan is this, We entered through the back door, disarmed the alarm, and loaded the jewelry into the car.

Security: Hey! You can't go there!

John: Be honest! We have guns!

Security: Don’t shoot! I have no weapons.

Second scene: ​

John: Damn it! In the end what happened?

Bert: I don’t know, the alarm must have gone off.

John: Damn it! Everyone, stay calm and we will find a solution.

Bill: What should I do? The police are coming!

John: We're using hostages as roadblocks. Everyone obeys my command.

Look at the plot, it’s pretty decent, and the tension is really high. I don’t know if it will be a big hit if it’s actually filmed.

The AI ​​Script Film Festival was released, and human experts were full of praise

In fact, there are related papers about the AI ​​playwright Dramatron developed by DeepMind The preprint was released as early as September this year, and DeepMind also promoted it on its blog.

Paper address:

​##​//m.sbmmt.com/link/177da57035b03c2eb2cbe9b872348d15​​

Four papers Among the authors, one of the publishing units is from Stanford University, but this research was conducted while he was at DeepMind.

Dramatron is a system that uses a large language model to generate scripts and screenplays through a "layered story generation" method.

Dramatron takes advantage of large language models to combine carefully designed cues and cue chains with structured generation to achieve long-distance coherence throughout the script. Compared with previous continuous text generation, Dramatron's script creation process can make the story more coherent.

Dramatron generates entire scripts, which can be tens of thousands of words long, from a user-supplied summary of a drama's main conflicts (called a "log line").

Based on the input log lines, Dramatron can generate the entire script, including the title, character list, storyline (i.e., a summary list of scenes with settings and beats), location descriptions, and dialogue.

Users can intervene and modify at any stage of layered generation, enter alternative generation, edit and rewrite output text, or continue text generation.

Dramatron's layered coherent story generation, starting with log lines, generates titles and characters.

In this way, the generated characters are used as prompts to generate a series of scene summaries in the plot. A description is then generated for each unique location. Finally, these elements are combined to generate dialogue for each scene. The arrows in the figure indicate how the generated text is used to build prompts for further generation of LLM text.

Our approach can be used with any LLM that accepts input prompts and then predicts which tokens will appear next.

At the International Frontier Theater Festival in Edmonton in August 2022, a series of scripts written by Dramatron according to this process will be screened.

To evaluate Dramatron's usability and capabilities, the researchers did not rely on crowdsourced annotations online and ratings from non-expert reviewers, but instead asked 15 experts to participate in a two-hour long User meetings, co-writing scripts with Dramatron.

Expert playwrights and screenwriters from the theater and film worlds provided feedback on Dramatron’s interactive co-creation process and provided insights into the creative outcomes artistic opinions and analysis.

Judging from the expert revision feedback provided by DeepMind, many of them are revisions to the details of scenes and dialogue wording, mainly Polish the narrative rhythm of scenes and dialogues, and make some modifications to avoid infringement.

The above figure shows the prompt setting instructions of the language model. The prompts generated by the user or Dramatron are connected to a prefix and equipped with a label. For different random seeds, several header outputs are produced.

What is the overall feedback rating from experts and professional editors? It can be seen that most expert reviewers still gave positive evaluations.

For example, the praise for Dramatron's interactive layered generation of lines includes words such as "useful", "unique", and "surprising".

In addition, many experts have spoken highly of the quality of the script written by AI. They believe that as long as the script is modified, it can really be filmed and released.

In addition, expert Dramatron expressed his expectation for more uses besides writing scripts. For example, it can be used as a heuristic tool for film and television screenwriters. Authors can also use lines and scripts generated by Dramatron to Draw ideas from scene conception.

In their view, AI provides materials and real authors write scripts. This model has great potential in the future.

Although in the eyes of human experts, scripts written by AI still have some unnatural flaws and problems. For example, the plot is too straightforward and easy to guess, the line style is a bit rigid, some lines are full of prejudice, sometimes the plot enters a logical loop and cannot be jumped out, etc.

However, with the treasure trove of large language models and the wisdom of playwrights and film and television workers around the world as training data sets, it is possible that future screenwriters of film and television dramas will be human or AI. , we may not be able to tell the difference in front of the big screen.

​References:

https://arxiv.org/abs/2209.14958

https://deepmind.github.io/dramatron/details. html

https://www.deepmind.com/publications/co-writing-screenplays-and-theatre-scripts-with-language-models-an-evaluation-by-industry-professionals

https://twitter.com/iamvladyashin/status/1601300031352745984?s=46&t=RzfB3tz8vKRsdX33EmaJmA

The above is the detailed content of Musk likes it! DeepMind's amazing AI screenwriter can generate tens of thousands of words of script in one sentence. For more information, please follow other related articles on the PHP Chinese website!