Sign In

Interview with Marcos: Game Asset Model Creator

In this interview, we speak with Marcos about his journey in the AI art space and his experience creating models for games.

When did you first get into the AI Art space? What attracted you?

Marcos got into the AI art space because of Etsy and was initially interested in logo and avatar design.

When did you start making models? What initially interested you in creating a model?

Marcos started making models on November 4th and was attracted to the idea of creating games. He began attracting an audience and customerbase through Twitter.

What was your process when you started?

When Marcos first started, he used Google Colabs to create his models.

What's your process now?

Now, Marcos works full-time as a model creator for a game studio. His process hasn't changed much since he started, but he has introduced heavier use of text encoder which allows him to specify how much to train the model. This has worked well for styles especially. With a 70-80% text encoder percentage resulting in a model that closely resembles the subject. For character creation, Marcos uses a text encoder of 55-65% for the first pass to allow for more creativity in the model. He also uses img2img to generate additional images for training and takes 100 steps per image. Marcos uses a HuggingFace/Private Dreambooth web interface for his model creation.

What challenges do you face when working with AI art models, and how do you overcome them?

Marcos faces the challenge of having to remake his model 80% of the time due to steps being too low or the text encoder not being aligned. He overcomes this by having good organization for his training images and sets clear goals for what the end result should look like, and by being thorough in his debugging process, which includes checking the text encoder percentage, steps, and dataset. He notes that having too much variety in the dataset can lead to straying too far from the goal.

Can you share some examples of your work and describe how you were able to achieve those results?

One example of Marcos' work is his ARMOR model. He first created a texture/style model and then applied it via img2img. He designed a basic armor outline and applied the textures, using the resulting images as the dataset. He then fine-tuned the model by slowly adding modes while creating the dataset.

Armor asset example

How do you see the field of AI art evolving over the next 6 months? Where do you think it'll be in a year?

Marcos is unsure of where the field of AI art will be in the next 6 months, but he thinks that in a few years, there will be the creation of usable 3D assets via text prompts and the possibility of a game where the story is generated based on character actions, similar to chatgpt.

In your opinion, what are some of the ethical considerations that need to be taken into account when creating AI art?

Marcos believes that artists would be well served to addapt this technology rather then be made obsolete by it. As an artist himself he understands the frustration that people can/will be able to create without having to learn the same skills he spent many years developing, but that ultimately this kind of power is a great thing. Artists need to look at the big picture; imagine how much this technology can speed up processes, and if all these untrained individuals are able to create such cool pieces then imagine what they, as trained artist, could accomplish.

How do you think AI art can impact society and culture?

Marcos believes that AI art has the potential to revolutionize the way that art is created and consumed. He thinks that traditional digital creators will have to adapt or risk being left behind, and that search engine optimization may become obsolete as chatbots like chatgpt can serve up customized results. He also wonders if it will make sense to create websites in the future when chatbots can scrape them and serve them up without ads. Despite these challenges, Marcos has an optimistic outlook, believing that we can adapt and overcome as we have done in the past.

What advice would you have for someone who is interested in creating AI art models?

Try it. Try, fail and try again until you get the hang of it.

For creating models on characters or subjects, where you may only have a few images, try to decribe character with text and generate images that are close to the character or subject, (uses midjourny for better results) then take those images to train a model.

If you could have any tool to make things better what would it be?

A tool that takes all the tweaking out of your hands. Steps, dataset, and text encoder would be set to optimal levels from the get go, and if not you would be alerted to them, or better yet it would do what it needs to make the setting and data optimal.

Are you an artist, model creator, or otherwise involved in this technology as it develops? Do you have some insights you like to share? Reach out to us at [email protected] so we can get in touch!