Frontend · Artificial Intelligence

A 3D Virtual Human Lands on the Desktop, Built Entirely by Talking to an AI

By 石小石Orz · Jul 25, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Client-side rendered 3D digital humans with sub-500ms latency remove the bandwidth and delay barriers that kept interactive avatars tethered to cloud streaming. Combined with AI coding tools that can scaffold an entire Electron app from a prompt, a solo developer can now ship a desktop virtual companion in hours instead of weeks.

Summary

A new desktop application puts a real-time, 3D-rendered virtual human directly on the desktop, built entirely through natural-language conversation with TRAE SOLO. The app packages the Mofa Nebula SDK inside an Electron shell, connecting to any LLM for streaming dialogue. The virtual human renders client-side with sub-500ms latency, supports voice interruption, and syncs expressions with speech.

The entire build—from Electron + Vue 3 project scaffolding to SDK integration, DeepSeek chat wiring, UI construction, and Windows installer packaging—was generated by describing requirements to TRAE SOLO. A pre-installed skill pack, `nebula-avatar-demo`, steered the AI toward correct SDK API calls, avoiding the trial-and-error loop common with less-guided code generation. When bugs surfaced, like an undraggable frameless window or a white screen after packaging, a single-sentence description triggered an automatic fix.

Mofa Nebula’s SDK uses a parameter-streaming approach for client-side rendering, sidestepping the high latency and bandwidth costs of cloud-streamed video digital humans. The platform positions itself as the missing “interaction layer” for LLMs and agents, giving AI a body, voice, and real-time expressiveness across terminals.

Takeaways

— TRAE SOLO generated the full Electron + Vue 3 + TypeScript project, including SDK integration, UI, and packaging config, from natural-language prompts alone.

— The Mofa Nebula SDK uses parameter streaming for client-side 3D rendering, achieving end-to-end response latency around 500ms.

— Installing the `nebula-avatar-demo` skill into TRAE SOLO prevented incorrect SDK API calls by steering code generation toward official best practices.

— Bugs like an undraggable frameless window or a post-packaging white screen were fixed by describing the symptom in a single sentence to TRAE SOLO.

— The final output is a single portable `.exe` file that runs without any environment setup, requiring only an SDK key to activate the 3D avatar.

— Registration with the invitation code `JM2A7C9LSW` grants 1,000 platform points, compared to 100 without it.

Conclusions

Guided code generation via installable skill packs marks a shift from generic AI coding toward domain-specific, reference-driven output that reduces API hallucination.

Client-side parameter streaming for digital humans eliminates the cloud rendering bottleneck, making real-time 3D avatars viable on consumer hardware without a GPU farm.

The project collapses three traditionally separate roles—3D rendering engineer, Electron developer, and LLM integration specialist—into a single prompt-driven workflow.

Concepts & terms

TRAE SOLO

An AI-powered development mode within the TRAE IDE that generates complete application code from natural-language descriptions, supporting full-stack and desktop app creation.

Mofa Nebula SDK (Embodied Interaction Intelligence)

A platform SDK that provides client-side 3D digital human rendering, voice synthesis, expression sync, and real-time interaction, acting as the 'body and expression' layer for LLMs and AI agents.

Parameter Streaming (Digital Humans)

A rendering approach where animation parameters (bone transforms, blend shapes) are sent to the client instead of video frames, enabling low-latency, locally rendered 3D avatars without cloud video streaming.

Skills (TRAE SOLO)

Installable extension packs that embed domain-specific code patterns and best practices into TRAE SOLO, guiding the AI to generate higher-quality, API-correct code for a particular SDK or framework.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗