Skip to main content
European Commission logo
Enterprise Europe Network

A Dutch company is looking for possibilities to train Lora/Dreambooth models at scale

Summary

Profile Type
Technology request
POD Reference
TRNL20231124014
Term of Validity
28 November 2023 - 27 November 2024
Company's Country
Netherlands
Type of partnership
Research and development cooperation agreement
Targeted Countries
All countries
Contact the EEN partner nearest to you for more information.
Find my local partner

General information

Short Summary
A Dutch creative production agency is looking for possibilities to train AI models at scale for a large dataset of ex. 1000 different items of clothing using imagery training methods like LoRA or Dreambooth to make it more automatic and efficient. Partners are sought in the frame of a research and development cooperation agreement. This request is part of an innovation challenge.
Full Description
The Dutch company is a purely digital operating brand that connects 8,550+ digital natives across one global team. They are united by a mission to shift industries forward and pave the path towards ambitious outcomes so their clients and people can realize their full potential for growth. Their unified model combines solutions in media, data, social, platforms, studio, experience, brand and technology services to help clients continuously reinvent themselves throughout increasingly rapid cycles of disruption. Their efforts to shape culture, build innovative technologies and unlock the future of growth have earned recognition from numerous esteemed panels.

Artificial Intelligence, image generation and model training have evolved over the past year in terms of capability and companies that use these tools will have significant advantages. For the Dutch companya this perfectly aligns with their vision of digital and craft. Some of their current projects use a Kohya repository specifically for training models based on images or styles of clients. The thing is, it’s slow and labor intensive.

Why is this a challenge?
The manual process of creating separate folders, captioning images, and resizing them in different programs poses several challenges and inefficiencies. Firstly, it is time-consuming and tedious, especially when dealing with a large number of images, ranging from 15 to 500. Secondly, the need to switch between different systems or tools for each task adds complexity and increases time. Also, training models can take up to 100 hours, require high GPU’s and you never know what kind of results you’re gonna get, it’s still highly experimental because of how new the Kohya system is.

With a lot of projected requests of projects/pitches utilizing image-based model training, there is a need for a more efficient and streamlined process that can save time and resources. They use an efficient web programme for batch image resizing and constantly look for updates and solutions online that could help them with efficiency, but it’s still a mix of different apps and a lot of manual labor.

The problem lies in the lack of a unified system that can automate the entire image training process. Having a single solution that can handle folder creation, image resizing, automated captioning, and giving different possible outcomes simplifies the workflow, saves time, and reduces the chances of errors.

Kohya is a fairly new system that is still in its experimental phase.
Challenging because of required GPU power and whilst training there are a lot of rules and parameters you have to know and follow to get good results.
You can’t preview the output, sometimes you have to wait a 100 hours for the training to end, only then you can test the model in another programme (ex. Stable Diffusion).
Manual labor, iteration and hours put into captioning, experimenting, testing and in using it in different programmes instead of one.

Partnership will be framed in research and development cooperation agreement. When fits the purpose also other agreements can be envisioned.

IMPORTANT NOTE: This technology request is part of an innovation challenge and is published on an open innovation platform from the 13th November 2023 and will close on the 7th of January 2024. If an organization does express interest in collaboration with this company before closing date, it will be guided through this open innovation platform and will be introduced to its moderator and the company’s experts. Mind that posts on this platform are not confidential. Next step is that the company will select the SMEs with whom they would like to cooperate in the development of the solution.
Stage of Development
Concept stage
Sustainable Development Goals
Not relevant

Partner Sought

Expected Role of a Partner
The company is looking for an innovative and unified solution that automates the image training process. By leveraging your expertise in AI, image processing, and software development, you can create a program or code that eliminates the need for manual folder creation, image resizing, and individual captioning. Your contribution will streamline the workflow, save time, improve efficiency and enable reliable model training.

What's in it for you?
Show off your skills in developing innovative solutions for automating image training, gaining recognition in the emerging tech industry
By making workflows more efficient, you contribute to the demand for sustainable AI tools, by default
It's new, exciting and creative and there is huge demand for solution building this into a product could lead into a lot of demand in use of our created service.

Which disciplines are needed?
Artificial Intelligence (AI) and Machine Learning (ML)
Software Development and Programming
Image Processing, Generative AI, Computer Vision
Type and Size of Partner
SME 50 - 249SME <=10SME 11-49
Type of partnership
Research and development cooperation agreement

Call details

Coordinator required
Yes
IPR description
We’re looking for an innovative and unified solution that automates the image training process. By leveraging your expertise in AI, image processing, and software development, you can create a program or code that eliminates the need for manual folder creation, image resizing, and individual captioning. Your contribution will streamline the workflow, save time, improve efficiency and enable reliable model training.

Dissemination

Technology keywords
01003020 - Building Automation Software01003006 - Computer Software
Market keywords
02006009 - Other computer services
Sector Groups Involved
Digital
Targeted countries
All countries