EA | Emmaparkervfx

EA Games AI Research Collaboration with SCADpro

Task: Our team was tasked with researching how one artistic style can be reflected continuously throughout varying concepts using AI to cut down on manual implementation.

My Contribution

Initial Experimentation with Meshy.ai

Test 01

This was the result. I have included this up close to show the errors within the model. From afar, the model looks great, but up-close, the knives are skewering the fingers.

Test 01

We wrote a brief description that was then used in a text to 3D feature.

Test 01

We began the experimentation by having one of the concept artists pitch a character concept. We settled on a gambling king pin inspired by the card faces.

Next Steps

Test 02

I wanted my next experiment to be with a simpler concept to visualize what works and what doesn't on a smaller scale.

Test 02

I prompted the AI to generate a model of a cowgirl with doll like features, detailing the clothing and hair style I wanted.

Test 02

This was the result.

Reflections

Based on this generation, I realized that the program generated a posed model making it virtually unusable. I was able to note the need to mention A/T posing in the future. Secondly, when describing the physical features of the character, I realized what the AI actually followed through on and what it didn't. More subjective descriptions like "big eyes" and "loose fitting top" were ultimately interpreted differently than I had imagined.

Notes for the future: be more intentional with descriptions

Implementing Changes and Exploring other Programs (Claude, ChatGPT, Nano Banana, Gemini)

Test 03

I tried to implement the notes I had from the initial generations. I was able to get a more accurate result in terms of posing posing.

Test 03

What I noticed was things like pants and shirt details struggled. For example, the shirt details including fringe and wrinkling look unnatural and illogical. The pants look like they are basically painted on. Stylization struggled to come through making the model look like an uncanny attempt at realism. Base clothes like under shirt and pants look like human anatomy that had been blurred.

Test 03

Some up-close examples of the details.

Test 04

I wanted to see what would happen if I used the same prompt repeatedly.

Test 04

Each result was different from the previous one. The program would sometimes add things that were associated with the words I prompted. Like cowboy translating to cowboy hat.

Test 04

If a specific style was not indicated it would generate a variety of styles ranging from realistic to cartoon. Even if a style was specified, the program would create variations of what it perceived that style to be.

Test 05

During this round of testing, I wanted to see what descriptive words were most effective since my previous attempts seemed to interpret words differently. I wanted to see how descriptive I could get.

Here I implemented the use of AI like Claude, ChatGPT, Gemini, and Nano Banana to polish my prompts to create a more effective description.

Even with the assistance, this model looks particularly bizarre. It is clear that my "long, full" hair description was interpreted in a hyperbolic fashion. The hair covers the entire upper half of the model front and back and the feet are twisted around.

Test 05

In this iteration, I made a point to describe the physicality of the model more in depth which clearly was reflected in the model. However, I found when more descriptive terms were added to expolain the body of a model, it would override any stylistic requests making it more anatomically correct.

Test 05

In this example, I broke down each aspect of the model giving it as many stylistic descriptions as possible. Most aspects of the character fit within one style which was helpful. I was able to use this prompt to inspire future generations. However, because quality is unpredicable, we didn't want to rely too heavily on this example.

Test 06

More prompt testing.

Test 06

More prompt testing.

Test 06

More prompt testing.

Reflections

Results are incredibly inconsistent. Even with detailed prompts, the output is oftentimes different enough from the previous attempt that the style does not carry across models. Moving forward, I wanted to see how we could better prompt the style of the models to get a clearer and more intentional result. Our goal is to push the software until it breaks.

Style Interpretation

Test 07

I was curious to see how concept art would aid in stylization. Ironically, the software struggled a bit more.

Test 07

We discovered that character turn arounds are not helpful when creating stylized models. Even when prompted differently, the program reads each angle as a separate character. This is what resulted in the bizarre shapes behind the central character.

Test 07

This is a shot showing more errors.

Test 08

Our next experiment was creating more descriptive style references to see how the results would vary. For example, this model used the same general prompt as the following two, but we changed the style the program should reference. This was a model inspired by Picasso's painterly style.

Test 08

This was another iteration of the same character inspired by Disney characters. We had described the character as being "mousy". The program took this literally.

Test 08

This character was inspired by the Seven Dwarfs. We also used Meshy.ai's animation feature which led to stretching and warping of the character.

Reflections

Style cannot be determined strictly based off prompt descriptions. Even when aided by concept art, results vary. Ideally, a proprietary software would be created developed by a collection of concept art. Style relies on thing other than character modeling so we wanted to continue the experiment with motion and texture mapping. This exploration would feed into the teams' later experiments with concept art and prompting.

Stylized Texturing Within Meshy.ai

Ashley Jones

Add a short bio for each team member. Make it brief and informative to keep visitors engaged.

Don Francis

Add a short bio for each team member. Make it brief and informative to keep visitors engaged.

Alexa Young

Add a short bio for each team member. Make it brief and informative to keep visitors engaged.

The Motion Team

Animation and Motion Exploration Using AI Software

Test 09

We tested out a ranch of programs to see which were able to create results that we were looking for. We prompted Nano Banana to generate a cartoonish walk cycle of a character inspired by Mario. It was only able to generate images.

Test 09

This was one of the tests we ran with Vizom. The results were of a basic character that shifted its weight from one leg to the other. There were some errors. Some iterations included the character standing still and swinging its arms. In others, the character would take a single step. In this example, the character walked into the edge of the screen while a third leg was generated in the left side of the screen.

Test 09

Here, we used Deevid to generate a video. Instead of creating a walk cycle, it generated a series of images played in sequence following the key words listed in the prompt.

Motion Designer, Move AI, and Motion Capture

I took a class working with motion capture and had the opportunity to work on a motion capture stage. Here, we recorded my movements and applied it to a couple of the AI generated characters made in our initial experiments. This is where I also experimented with Move AI and motion designer.

Art Director

Ashley Jones

Add a short bio for each team member. Make it brief and informative to keep visitors engaged.

Tech Lead

Don Francis

Add a short bio for each team member. Make it brief and informative to keep visitors engaged.

Product Manager

Alexa Young

Add a short bio for each team member. Make it brief and informative to keep visitors engaged.

Final Conclusions

My hypothesis is that with the help of motion capture and collected animation data we could create a merge between the intended style, and the existing movement to create a style baseline.

For future research I would love to explore how animation curves utilizing exaggerated movement could be compared to more realistic animation curves. From here, we could average the changed that a point on the existing curve would need to move to emulate that type of style. By using a software based around the style of motion designer, we could potentially blend data together to warp said curves.

EA Games AI Research Collaboration with SCADpro Task: Our team was tasked with researching how one artistic style can be reflected continuously throughout varying concepts using AI to cut down on manual implementation.

My Contribution

Initial Experimentation with Meshy.ai

Next Steps

Reflections

Implementing Changes and Exploring other Programs (Claude, ChatGPT, Nano Banana, Gemini)

Reflections

Style Interpretation

Reflections

Stylized Texturing Within Meshy.ai

The Motion Team

Animation and Motion Exploration Using AI Software

Motion Designer, Move AI, and Motion Capture

Final Conclusions

EA Games AI Research Collaboration with SCADpro

Task: Our team was tasked with researching how one artistic style can be reflected continuously throughout varying concepts using AI to cut down on manual implementation.