AI Avatar Generation Service "DigiSelf"

“DigiSelf”, offered by EmbodyMe Inc., is a service that generates AI avatar videos of a quality indistinguishable from real humans, from just a single photo and a text input. Beyond video generation, it also enables two-way communication with AI avatars through real-time generation technology.

Since May 2025, we have served as DigiSelf’s development partner, building its cloud infrastructure, the execution platform for its generative AI models, the API backend, and the web frontend. This article introduces the background of the project and the technical approaches we took, together with comments from our interview with EmbodyMe CEO Issay Yoshida.

EmbodyMe and DigiSelf

Since its founding in 2016, EmbodyMe has researched and developed foundational generative AI technology in the field of video. Its video generation AI holds a globally competitive edge in real-time processing performance, with accolades including the NVIDIA Inception Award and the Microsoft Innovation Award, as well as multiple international patents. The xpression series, including its predecessor “xpression chat”, a consumer app for conversing with AI avatars, has 95% of its users overseas and reached #1 on Product Hunt.

Meanwhile, most AI video generation services in Japan rely on AI models developed overseas, and their Japanese-language support is often insufficient. DigiSelf was launched to bring this technology to Japanese companies as a service that relentlessly pursues Japanese-language quality. Where xpression chat was “consumer-facing and overseas”, DigiSelf is “business-facing and, to begin with, domestic”. Built on the same technology, it is also a bold shift in business target.

Project Background

The relationship between our CEO Mizutori and Mr. Yoshida goes back to South by Southwest in Texas. At the time, Mr. Yoshida was staffing his previous company’s booth, while Mizutori was helping out at the Sony booth at a friend’s request. The two booths happened to be next to each other, and they struck up a conversation. Later, they crossed paths again by chance at “TechLabPAAK”, the open innovation space Recruit ran in Shibuya. It offered selected startups a six-month residency, and EmbodyMe and our company were residents at the same time. A few years ago, Mr. Yoshida then commissioned us to develop xpression chat. On that project, EmbodyMe developed the complex avatar rendering library and the backend, while we were responsible for the UI/UX and application logic of the iOS and Android apps.

Then, in 2025, he reached out to us again for the development of DigiSelf.

We simply didn’t have the staff to build out a service this large. When we decided to entrust the service layer to an outside partner, Goldrush, who we had worked with all along, was the first to come to mind. On the technical side there was already a sense of trust. Having worked together for so long, we knew their technical capabilities, so we could hand it over with complete confidence, as a given. (Mr. Yoshida)

Our Scope of Work

Avatar model generation and avatar video generation, the “intelligence” of DigiSelf, are proprietary technologies that EmbodyMe has spent years researching and developing. Our role was to turn that intelligence into a working service.

• Design and construction of the cloud infrastructure
• The execution environment for the avatar generation and video generation models, and the mechanism for on-demand, scalable execution
• API backend
• Web frontend

Development began in May 2025, and the system went live at the end of August as the first delivery. Prospective customers started using it from that point, a pre-release to selected companies followed in October, and the service officially launched on March 30, 2026.

The Technical Challenge: Keeping GPU Costs Down

When asked what was most difficult throughout the development, Mr. Yoshida named GPU costs without hesitation.

Avatar generation runs on GPUs, so it all comes down to how far you can keep GPU costs down. Generation speed and cost are two sides of the same coin: the longer generation takes, the more it costs. So the biggest thing was precisely the job-dispatching execution mechanism they built for us. (Mr. Yoshida)

Generating avatars and videos requires high-performance GPU servers, but GPU servers are extremely expensive, and keeping them running around the clock means paying for idle time during low-usage hours. During development, proof-of-concept trials, and the pre-release stage in particular, usage patterns are hard to predict, so this standby cost weighs heavily. We therefore designed an on-demand execution mechanism that scales GPU servers down to zero between generation requests and spins up only as many servers as needed when requests arrive. We focused on controlling server start-ups according to request conditions so that unnecessary launches are kept to a minimum.

We also asked about operations after launch.

The scaling mechanism for the AI avatars is working really well. We designed it together, and it runs on a very sound design. We’re grateful for that. (Mr. Yoshida)

Communication During Development

When asked about proposals during development that left an impression, this was the answer.

They often proposed and implemented things we hadn’t even thought of, which was a huge help. Take the watermark, for example. Without a drop shadow, it blends into white and disappears on videos with white backgrounds. They suggested adding a shadow there. It was something we hadn’t considered, so we were very grateful. (Mr. Yoshida)

Details like the watermark overlaid on generated videos are the kind of thing you don’t notice while defining requirements and specifications. Proposing on the spot what we notice while building something that actually works, and carrying it through to implementation when the client agrees it is a good idea, is something we value in our contract development work.

Expanding Use Cases

DigiSelf’s video generation is intended for use in sales, recruiting, training, marketing, PR, social media, and more. When content needs to change, only the relevant parts need to be swapped out, with no reshooting required, which dramatically reduces video production costs. Real-time generation can be applied to customer service, sales, customer support, interviews, digital signage, and more. Because a single person can serve multiple customers at once, it also helps address labor shortages.

One thing that stood out in the interview was the demand for “CEO avatars”. Regardless of company size, there is considerable demand from companies that want to turn their CEO into an avatar and deliver messages to employees and external audiences. In place of a busy executive who cannot find the time, their own avatar can deliver messages, even daily. Employees know they are watching an avatar, yet hearing a message in the person’s own voice lands completely differently from reading it as text. And once real-time dialogue becomes possible, it becomes a different experience altogether.

Evaluation as a Partner

Finally, we asked what kind of company he would recommend us to.

We’ve worked with large companies before, but they often lack flexibility. Goldrush listens to even the smallest requests and thinks them through with us, and honestly, even when we’ve asked for a bit too much, they’ve accommodated us to a reasonable extent. They’re agile, we can work very closely together, and on top of that, their technical capabilities are high. I think that’s what makes them so good. (Mr. Yoshida)

DigiSelf has continued to evolve since its release, and we have been in charge of developing its feature expansions.

When launching a generative AI service, it is not only the performance of the model itself but also the design of an execution platform that keeps GPU costs within a realistic range that determines whether the business is viable. If you are considering launching an AI service or building infrastructure for GPU-based workloads, please feel free to reach out to us.

Contact form: goldrushcomputing.com/contact