In what ways can Depix Design Lab be deployed at our facility?

1. Local installation on Windows Workstation (per user) - accessed via web browser on the desktop.

2. Local deployment on Windows Server/Linux server (in on premises server room) - accessed via browser on the internal networks

3. Local deployment of certified Depix / Hewlett Packard “appliance” - hardware delivered with software installed and certified by HP and access via local internal network by browser.

4. External deployment on your cloud provider accessed via internal web browser

5. External deployment on Depix designated cloud provider with Depix authentication (we have begun the process of SOC2 compliance).

Which deployment is best for my company?

We can deploy in a number of ways as indicated above. Right now our recommendation would be to do an on premises installation on Windows Server or Linux that resides on your internal local network that can be accessed remotely by your users via web browser. As our cloud implementation becomes more mat

What web browsers are supported?

Design Lab is supported on Mac and PC with Google Chrome, Firefox, Microsoft Edge. Safari is not supported at this time.

What operating systems are supported?

Operating System Windows: 10, 11, Server 2022, Server 2019

Linux: Ubuntu Server 22.04 LTS, Amazon Linux 2

If we increase the RAM or GPU VRAM does that affect the speed of rendered

results? Are there any support limits?

RAM does not have a significant impact on the speed of rendering. Increasing the VRAM in the GPU can speed up the generation process. Additionally, the type of GPU also plays a role in the speed of generation. Several key points should be considered to determine the performance of the GPU: VRAM Size, Clock Speeds, Memory Bandwidth. When multiple users are using the application concurrently, expanding the number of GPUs speeds up the system by distributing tasks across parallel processing units. This parallelization enables simultaneous processing of tasks, resulting in faster overall system performance.

We don’t have any support limits. We will work with you to ensure implementation is satisfactory for any size deployment.

Does Design Lab run on a Mac?

Local installation on a Macintosh is not supported right now. However, you can access the Design Lab that is on your network using Chrome, Firefox or Edge, (Safari is not currently supported)

What is the recommended hardware configuration for Depix Design Lab?

Since Design Lab uses an AI inference engine, the latest Nvidia graphics cards supporting AI are best. These include those listed below.

Example recommended graphics cards and their VRAM for Depix Design Lab:

A10G 24GB VRAM 

A100 80GB VRAM 

RTX6000 24GB VRAM  

RTX6000 ADA 48GB VRAM  

We strongly recommend a dual GPU system with each card having a minimum of 16 GB VRAM. Recommended is Dual RTX6000 ADA with fast CPU/32 GB Ram minimum and multiple terabytes of SSD disk storage.

What’s needed to support 100 users?

The AI image generation process used by Depix requires a max of 16 GB VRAM per generation. Adding a second GPU means that training processes will be distributed to the second card, leaving the first for image generation.

Depix is in the process of testing dual and quad GPU machines and optimizing for multi-users. We will soon have more information on the requirements for supporting 100 users. The number of machines will be determined by the number of simultaneous users and the tolerance for waiting for inference results. We are working on enabling multiple processes on a single gpu and general performance enhancements that will be released in the next few weeks.

How can AI models be developed on historical data?

Enhancing the base foundational model with your data is highly recommended. By gathering images of particular models, years, designers, or other categories, images need to be sorted, categorized, and labeled with as much detail in the metadata as possible in a database that can be exported and imported into Design Lab for training. For large datasets (>1000 images), we recommend working directly with you on a consulting basis for the optimization of the process, as there will be some special considerations based on your needs. Data can be categorized in a spreadsheet. We have performed several tests on interior and exterior data sets in the 3500 image set range and found the improvement to the foundational model is significant.

What types/categories of images can be used for tuning/training the AI models?

1. Clinic images

2. Project milestone images

3. Design sketches (interior/exterior)

4. Individual designers work

5. Concept show cars 

6. Images of elements with particular brand characteristics

7. Product new release marketing images

8. Collection of images representing brand characteristics

9. New competitor car introductions at a particular car show

10. Production cars by year / model

11. Production interiors

12. Influential shapes

13. Influential sculptures

14. Parts like headlamps, tail lamps, wheels, interior components.

15. Motorsports

16. Historical cars

17. Fashion/Clothing

How many images should I use in a tuning/training set?

Tuning the base model can be just a single image. However, this as influence is really no different than just using a single image as influence in the generate tab. 

Typically 5-30 images are necessary for an adequate model tuning. The training area is set up to accept anywhere from 1 to 1000 images. If more than 1000 images need to be in a data set, there are different parameters and probably some trial and error necessary to make sure the settings are correct. We can work with your team on a consulting basis to enable larger training sets. At Depix we have trained models with more than 500K images.

Do I need to label the images in the tuning set?

In the example above there are three different car models. If you want to bring influence from just one particular model in this tuning set of three models, you need to manually label the data so the system understands when you use a particular model name in a prompt to pull from that part of the AI.  The more detail you add to each image the better, but labels are not required for smaller tuning sets.

A keyword can be added to the tuned model that can be used to call heavy influence from the data set in an image generation. A key word needs to be a word not found in the English language. A good one to use is dpx_”tuning name”. When the key word is used in the prompt and the model is selected it gets explicit influence from the training that will be evident in the images generated.

What’s an ideal pathway to get metadata into the training set?

There is a large variety of meta data that might be stored with an image. It might include camera information like f stop, location, date, model, model variant, model number etc. We can automate any kind of format to make a connection between images with metadata and the training form. We would have to work with you to determine the best path, but it can be automated. Usually this type of data can be stored/exported in a spreadsheet.

The more information that can be included for manual labeling, the better.

Can I add images to an existing tuned model?.

Yes, if you want to add additional images or delete images from a tuned model, you simply open the set of images add/delete and hit retrain. 

Can I blend two tuned models together?

Yes, you can blend two models together with a weight slider.

What are the benefits of a cloud implementation?

• Initial Costs: Lower initial costs with no need for hardware or infrastructure investments.

• Operational Costs: Recurring subscription fees based on usage, but reduced costs for maintenance and staffing.

• Easy upgrades as only one machine needs to be installed.

• Data Security: Dependence on the provider's security measures, though reputable providers offer strong security protocols.

• Depix has begun the process of adhering to the SOC2 type 1 compliance and we understand that compliance to this standard is very important.

What are the benefits of onsite implementation?

• Data Security: Data remains within the organization's control, potentially offering higher security.

• Compliance: Easier to ensure compliance as all software is local and no connection to the internet beyond pulling the initial installation package is required.

• Responsibility: The organization is responsible for maintaining and updating the software.

• Updates may require downtime, which needs to be managed. Depix runs on three week sprints so we have updates every three weeks available but based on the content of the update, your organization might want to have a slower update cadence. We recommend it every six weeks.

Onsite vs. cloud?

• The choice between onsite and cloud-based implementation depends on the organization's specific needs, including cost considerations, scalability requirements, control and customization needs, security concerns, and maintenance capabilities

What is SOC 2 compliance for third party cloud implementation?

• SOC 2 is a voluntary cybersecurity compliance framework developed by the American Institute of CPAs (AICPA) for service organizations that specifies how organizations should handle customer data. The standard covers five pillars, called Trust Services Criteria (TSC): security, availability, processing integrity, confidentiality, and privacy.

• SOC 2 compliance is part of the American Institute of CPAs’ Service Organization Control reporting platform. Its intent is to ensure the safety and privacy of your customers’ data, that the company will comply with regulations, and that it has the processes in place to mitigate risk.

• SOC 2 is not a prescriptive list of controls, tools, or processes. Rather, it cites the criteria required to maintain robust information security, allowing each company to adopt the practices and processes relevant to their own objectives and operations.

The five trust services criteria are detailed below:

1. Security refers to the protection of information and systems from unauthorized access. This may be through the use of IT security infrastructures such as firewalls, two-factor authentication, and other measures to keep your data safe from unauthorized access.

2. Availability is whether the infrastructure, software, or information is maintained and has controls for operation, monitoring, and maintenance. This criteria also gauges whether your company maintains minimal acceptable network performance levels and assesses and mitigates potential external threats.

3. Processing integrity ensures that systems perform their functions as intended and are free from error, delay, omission, and unauthorized or inadvertent manipulation. This means that data processing operations work as they should and are authorized, complete, and accurate.

4. Confidentiality addresses the company’s ability to protect data that should be restricted to a specified set of persons or organizations. This includes client data intended only for company personnel, confidential company information such as business plans or intellectual property, or any other information required to be protected by law, regulations, contracts, or agreements.

5. Privacy speaks to an organization’s ability to safeguard personally identifiable information from unauthorized access. This information generally takes the form of name, social security, or address information or other identifiers such as race, ethnicity, or health information.