Midjourney Surprise: New Search about making LLMS write more creative

Midjourney It is known as one of the leading AI’s photo generators – with nearly 20 million users on Discord. According to the third party trackingAnd it is assumed that more on this site on its website – but its ambitions began to expand.

after News in late summer 2024 It was building its own computing and artificial intelligence devices. The company issued a new research paper this week alongside automated learning experts at New York University (NYU) to train the large text -based language models such as Meta Open Source Lama and Mistral Source Models with a more creative name.

Cooperation, documented in a New research paper It was published on Code Code Community Face face, introducing two new technologies – improving diverse direct preference (DDPO) and improving various preferences (DorPo) – designed to expand possible outputs while maintaining cohesion and reading capacity.

For the company that is famous for the AI’s generation models, the new Midjourney approach to re -thinking about creativity in the text -based LLMS shows that it does not limit his ambitions in the images, and that the image may actually be equal to a thousand words.

Could LLM be from Midjourney Native or a difficult version of LLM in the cards from the start -up starting up to boot? I arrived at the founder of Midjourney David Holz, but I haven’t heard yet.

Regardless of the Midjourney LLM width at the first end, the effects of its new research go beyond academic exercises and can be used to help fuel a new wave of LLM training between AI teams for institutions, product developers and content creators who are looking to improve the text created in artificial intelligence.

He also explains that despite the modern interest and investment between the service providers of the Models of Amnesty International in the models of new multimedia and logic, there is still a lot of the remaining juice to be pressured, cognitive and performance, from the classic LLMS that focuses on text.

The problem: The writing created from artificial intelligence collapses about homogeneous outputs

In fields such as the fact -standing Q&A or aid in coding, LLMS is expected to create the best response.

However, creative writing is open by nature, which means that there are many correct responses to demand one.

For an example made by Midjourney researchers, they gave a claim like “Write a story about a dog on the moon”LLM can explore multiple various tracks such as:

• He left a pet dog in an astronaut by mistake after the moon’s mission.
• A dog finds itself in the colony of future dogs.
• A dog with a way befriends a strange type.

Despite this range of possibilities, LLMS that has been seized in instructions is often converged on similar lines and topics. This happens because:

1. Post -training technologies give user preferences over originality, which enhances popular but frequent responses.
2. Adjustment of the instructions often leads to contrast, which makes models prefer “safe” responses over unique responses.
3. The current diversity enhancement techniques (such as temperature control) only work at the time of reasoning, instead of bread in the process of learning the model.

This leads to homogeneous stories, as creative writing created from artificial intelligence feels repetition and lacks surprise or depth.

The solution: modifying post -training methods to determine the priorities of diversity

To overcome these restrictions, researchers DDPO and DorPo presented two extensions for current preferences. The basic innovation in these methods is the use of deviation – a measure of how different the response of others – to direct training.

Here is how to work:

1. During training, the model is given the compromise and possible multiple responses.
2. Each response to others is compared to the same claim, and the degree of deviation is calculated.
3. Rare but high -quality responses are highly weight in training, and the model is encouraged to learn from various examples.

By integrating deviation in improving direct preference (DPO) and improving the preferences (Orpo), the model learns to produce high -quality but more varied responses.

This method guarantees that the stories created of artificial intelligence do not converge with a single predictive structure, but instead explore a wide range of characters, settings and topics-completely as a human writer.

What researchers in Midjourney did to achieve this

The study included LLMS training on creative writing tasks using a set of data from Subreddit R/Trintsprompts, a Reddit community where users publish claims and respond with short stories.

The researchers used two basic models to train them:

• Meta’s Llama-3.1-8B ((Model 8 billion of the Llama 3 series).
• Mistral-7B-V0.3 ((Form of 7 billion teacher from Mistral Ai).

Then they took these models through the following operations:

1. Service is subject to supervision (SFT): Models were first adjusted using Lora (low -ranking adaptation) to efficiently adjust parameters.
2. Improving preference:
   • DPO and Orpo have been used as the foundation lines– These standard methods focus on improving response quality based on user preference signals.
   • Then DDPO and DorPo were appliedProviding a penalty shootout to encourage more unique responses.
3. evaluation:
   • Automatic evaluation: semantic diversity and stylistic measured using inclusion -based techniques.
   • Human evaluation: Judges evaluated whether the outputs were varied and shared compared to GPT-4O and Claude 3.5.

Main Training Results:

• DDPO greatly outperformed the standard DPO In terms of diversity of directing while maintaining quality.
• Llama-3.1-8B has achieved the best balance Of quality, diversity, and production of responses More varied GPT-4O While maintaining cohesion.
• When the size of the data set was reducedDDPO models still maintain diversity, although they require a certain number of various training samples to be completely effective.

The effects of institutions: What does it mean for those who use artificial intelligence to produce creative responses – as is the case in marketing texts, storytelling of companies, text programming/television/video?

For artificial intelligence teams that manage the deployment of LLM, enhancing the diversity of directing while maintaining quality is an important challenge. These results have significant effects on organizations that depend on the content created by artificial intelligence in applications such as:

• Artificial intelligence conversation and chatbots (Ensuring various and attractive responses).
• Content marketing tools and storytelling (Preventing a frequent copy of artificial intelligence).
• Game development and narration design (Create a diverse dialogue and branching stories).

For professionals responsible for refining and publishing models in preparing the institution, this research provides:

• A new approach to LLM after training enhances creativity without sacrificing quality.
• A practical alternative to controlling the diversity of reasoning time (such as temperature adjustments) by combining diversity in the learning process itself.
• The ability to develop more attractive applications of artificial intelligence, from writing tools with the help of AI to virtual assistants that can adapt their dynamic responses.

For those who deal with the format of the artificial intelligence model and automation, this research is highlighted:

• The importance of seizure models in the training phase, which reduces the need for post -treatment modifications upon publication.
• A way to introduce adaptive stories novel in AI-drives AI applications, ensuring contrast while maintaining high content quality.
• A way to make LLM more human -like outputs, which is crucial for applications that require interactive stories, customer sharing, or dynamic content creation.

The future of the creative projects generated by artificial intelligence appears bright

The success of DDPO and DorPo explains that LLMS training with diversity -focused goals can result in significant improvements in creative writing. Some ideas include:

1. Merging deviation -based learning into the Models of Amnesty International for the Foundation To enhance the diversity of response in applications facing customers.
2. Explore how these methods are applied to other obstetric tasksLike poetry with artificial intelligence, script writing, or telling stories.
3. Development of hybrid training approach This balance Various capabilities and follow -up of instructions For artificial intelligence aides.

For those wishing to implement these technologies, researchers are planning to provide their symbol of the public to this Gaytap warehouse

Whether you are installing LLMS for business applications or improving artificial intelligence coordination on a large scale, this study provides implemented visions on how models are more dynamic, attractive and respond to creative tasks.

By adopting these technologies, artificial intelligence difference can exceed solid solid outputs – building artificial intelligence systems that are not only smart but also really fictional.