Midjourney’s surprise: new research on making LLMs write more creatively

Midjourney The best is known as one of the AI ​​image generators – with about 20 million users in his hatred channel, According to third-party viewersAnd it is likely to move more on the website – but its ambitions begin to expand.

Chase News at the end of 2024 This week, the open source LLAMA (LLS), Open Source LLAM (LLS), Open Source LLAM (LLS), this week, launched a new research paper with a new educational specialists to write more creativity.

Cooperation, documented New research paper The face of the AI ​​code presents the public, presents two new technical equipment – diversified direct selection optimization (DDPO) and diversified liberation (Dorpo) – Dorpo) – designed to expand possible exits while maintaining compliance and readings.

The new approach of the work in Midjourney’s text-based LLMs for a company that is best known to create a diffusion AI image, and a picture cannot really be worth a thousand words.

Can an existing LLM’s Midjourney-native LLM or a delicate adjustable version of a delicate adjustable in the cards from the boot start? I reached the midjourney founder David Holz, but I still heard back.

First of all, the results of the Midjourney LLM, the results of their new research can be used to help the new wave of a new LLM training between the enterprise, product development and content creators that develop the AI.

In addition, despite the latest interest and investment among AI model providers in new multimodal and meditational models, the classic transformer-based, text-directed LLMs are compressed, cognitive and performance wise are still left.

Problem: Ai created writing collapses around homogeneous speeches

In fact, such as Q & A or coding help, the LLMS is expected to create a best answer.

However, the creative writing is clearly open due to the presence of reliable answers with a single request.

Was desired for an instance given by Midjourney researchers “Write a story about a dog a month”Can explore as many different ways such as LLM:

• Astronaut’s pet dog was accidentally left behind after the mission of Aysar.
• A dog that finds himself in a futuristic canine cosmic colony.
• A closed dog with a strangely type of alien.

Despite the range of these opportunities, instruction-adjustable LLMs often combine similar stories and topics. This is because:

1. Post-training techniques prefer the user, strengthen popular but repetitive answers and prefer the originality.
2. Instructional regulation often smooths variability, provides “secure” responses on the unique ones of models.
3. Existing variety-incentive techniques (such as temperature tuning) are only in a fruitless time instead of cooking the model’s learning process.

This creates creative writing causes a homogeneous story that is not repetitive and no surprise or depth.

Solution: Changing post-training methods for prioritization of diversity

To relieve these restrictions, researchers presented DDPO and Dorpo, two extensions from DDPO and Dorpo, existing preferred optimization methods. This approaches are the use of basic innovation deviation – a measurement to measure how much the answer is different from others.

Here’s how it works:

1. During the training, the model in the model and numerous answers are given.
2. Each answer is compared to others for the same request and calculated a deviation account.
3. Rare, but high quality answers are more common in training, pleasing the model to learn from different examples.

Direct selection optimization (DPO) and the proportional optimization (OPO), the model learns to produce high quality, but more different answers.

This method ensures that the AI ​​has not become a predicted structure of the originating stories, vice versa provides more spacious characters, parameters and threads as a human writer.

Midjourney researchers did to achieve this

Research, SubredDit R / Canyprompts, teachers engaged in training LLMs using REDDIT Community, which users sent and responding with short stories

Researchers used two main models for exercise:

• Meta’s Llama-3.1-8b (8 billion parameter model from the yol 3 series).
• Mistral-7B-V0.3 (Mistral AI is a 7 billion-billion parameter model).

Then these models had the following processes:

1. Controlled delicate adjustment (SFT): Models were first beautifully used using LORA (low-ranking adaptation) to adjust the settings effectively.
2. Preference optimization:
   • Was used as DPO and Orpo Baselin-This standard methods are aimed at improving the quality of response based on user-choice signals.
   • DDPO and DORPO was then appliedTo apply deviation-based weight to promote more unique answers.
3. Assessment:
   • Automatic Rating: Semantic and stylistic diversity is measured using practice-based techniques.
   • Man’s assessment: judges assessed whether the results are different and attractive compared to the GPT-4O and Claude 3.5.

Basic Training Findings:

• DDPO is a significantly superior standard dpo in terms of exit diversity while maintaining quality.
• Llama-3.1-8b achieved the best balance with DDPO Quality and diversity, answers answers Is more diverse than GPT-4O while maintaining matching.
• When DataSet size decreasesDDPO models still diversity, although various samples demanded the full effectiveness of various exercise, they continued.

Effactivity: What does the AI ​​mean for those who use the creative answers – for example, are marketing corporate story, corporate story and films / TV / video game scripts?

Managing LLM placement for AI teams is a critical problem by increasing the diversity of exit while maintaining quality. These findings have a significant impact on organizations that trust the contents of AI in apps:

• Talk AI and Chatbots (varied and attractive answers).
• Content marketing and story tools (Prevent repetitive AI caused copy).
• Game development and narrative design (Create different dialogue and dampial stories).

This research provides this research for specialists who are responsible for subtle regulation and models in the conditions of an enterprise:

• A new approach to post-trained training in LLM, which increases creativity without sacrificing quality.
• Diversity regulation for the establishment of diversity by combining diversity in the learning process (such as temperature regulation) is a practical alternative.
• The potential to develop the potential to develop more attractive AI applications to virtual assistants who can adapt their answers to aI ai-auxiliary writing.

This research points for the AI ​​model orcervice and automation:

• The importance of tuning models in the training phase, reducing the need for post-processing adjustments.
• A way to provide variability, ensuring variability, the adaphant story story when holding the quality of content high.
• It is very important for applications that require a method, interactive story, customer tab or dynamic content to explain the more people in LLM.

The future of creative projects created by AI looks shiny

DDPO and Dorpo demonstrates the success of the Diversity LLMs, which are diversity-oriented purposes, can develop significant in the creative writing. Some ideas include:

1. Access to AI models to deviant-based learning facility Increase the diversity of response in applications facing client.
2. Explore how these methods apply to other generation positionsFor example, a Poetry, Scenario or game story to AI-power.
3. To develop hybrid training approaches that equilibrium Diversity and instructions-The following opportunities For AI assistants.

For those who want to apply these methods, researchers plan to present their codes to the public GitHub Depot

The delicate adjustment for business applications provides more dynamic, attractive, attractive and attractive, attractive, attractive and attractive, attractive, attractive and attractive, attractive and attractive, attractive and attractive, attractive, attractive, attractive and attractive, attractive, attractive, attractive, attractive, attractive, attractive, attractive, attractive, attractive, attractive and attractive, attractive, attractive, attractive, attractive and attractive, attractive, attractive, attractive and attractive, attractive, attractive, attractive and attractive, attractive, attractive, attractive and attractive, attractive, attractive, attractive and attractive, attractive and attractive, attractive, attractive, attractive, attractive and attractive, attractive and attractive.

By accepting these methods, the AI ​​teams are not only intelligent, but also a really imaginative, the formulaic exit and construction-building can go beyond the AI ​​systems.