← Testing dashboardTrace public implementations, prepare compatible inputs and measure actual outputs. Preparing a job does not upload files or start paid compute.
Our architecture
Script / uploaded audio → language-specific voice → engine adapter → GPU job → completed MP4 → facial diagnostics + human lip-sync review → measured cost.
Live conversation needs a separate streaming path: microphone → speech recognition → knowledge / AI response → streaming voice → real-time avatar. Batch video generation speed does not prove live-call readiness.
DuiX Avatar
Existing source video + new audio. The upstream application creates a queued job, calls synthesis, polls status and records the output path.
Read upstream job flowModel licensing and container/GPU setup still need deployment review. Source photo alone is not a video template.
LongCat Avatar 1.5
Reference photo + audio + descriptive prompt. Official recipe uses distilled inference and supports INT8. Published example uses two GPU processes; our 3090 compatibility is unverified.
Read official inference recipesPaid engines
HeyGen / Simli can be separate service adapters. Their public APIs can be integrated; their private model source is not recoverable from videos. No paid service is connected here.
HeyGen API documentationSimli documentation
Prepare a real engine input
Enter paths as seen on the GPU host/container. This page cannot confirm remote files exist. Use the same spoken audio across engines for a fair test.
No job prepared.
Download DuiX submit/status adapter · Runs on your GPU host. Default is dry-run; --submit starts the job. No automatic file transfer.
Measured render cost
Enter observed timings, not advertised performance. Includes allocated startup/idle time; excludes storage, voice, AI and tax.
No measured cost yet.
Review completed renders and facial readings →