Show HN: LemonSlice – Upgrade your voice agents to real-time video

Hacker News Discussion ↗

Real-time, AI-generated video avatars are emerging as a new interface layer for voice agents, enabling applications from training simulations and language tutoring to “AI coworkers” that can record Loom-style explainer videos. Commenters highlight LemonSlice’s technical leap—a 20B-parameter diffusion transformer streaming interactive video at ~20fps on a single GPU—and its flexible integrations with existing LLM, TTS, and voice platforms. Alongside enthusiasm about product potential and speed, participants raise concerns about audio quality, pricing clarity, and the societal impact of photorealistic agents in contexts like hiring, call centers, and human–AI trust.

Use Cases and Enthusiasm

  • Many commenters find the real-time video agents unusually impressive, describing “mind blown” reactions and extended tinkering.
  • Popular imagined uses:
    • Turning coding/chat agents into more “employee-like” coworkers that record Loom-style walkthroughs.
    • Roleplay-based training (nurses triaging patients, SDRs practicing sales calls).
    • Language tutoring, customer support, and website onboarding.
  • Some users already built demos (e.g., a golden retriever tutor) and report a strong “computer has come alive” feeling.

Architecture, Integrations, and Controls

  • LemonSlice is positioned as a video “avatar layer” on top of arbitrary voice agents.
    • API takes text and streams back synchronized video.
    • LiveKit integration allows plugging in OpenAI realtime, other STT/LLM/TTS stacks, or future S2S providers.
    • Hosted option currently partners with ElevenLabs; default LLM in their own stack is Qwen.
  • Users can influence avatar motion and emotion via text prompts; finer-grained motion control via API is planned.
  • Background motion is also prompt-controlled; better hand-motion control is in training.

Quality, Latency, and UX Feedback

  • Praise for A/V sync and responsiveness overall, but several issues noted:
    • Low resolution/FPS, inconsistent lip-sync, and “cheap mic” audio feel for some avatars.
    • Latency is noticeable, especially vs NVIDIA Personaplex; speed is a stated main focus area.
    • STT‑LLM‑TTS limits nuanced speech/pronunciation feedback (e.g., Spanish dialect practice); S2S is desired but currently too slow in tests.
    • Occasional visual “hallucinations” (e.g., pseudo‑Chinese subtitles).
  • UI confusions: 10s GPU spin-up looks like processing delay; demo video defaulting to 1.5x; privacy page unreadable in dark mode (quickly fixed); some mobile iOS issues (details unclear).

Pricing, Product Model, and Openness

  • Confusion around pricing: difference between “Video Agents” (interactive calls) and “Creative Studio” (downloadable clips) needed explicit clarification.
  • Real-time calls are fully streamed; there’s no native “record and replay exact answer later” feature.
  • Core model is a 20B-parameter diffusion transformer running ~20fps on a single Hopper GPU. Team expects similar approaches to be widely copied; they see substantial “low-hanging fruit” in real-time DiT optimization.
  • Open-weights release is under consideration; concerns are support overhead, not just customer cannibalization. No concrete commitment yet.
  • IP protection, profitability, and business metrics are asked about but not substantively answered in-thread (status unclear).

Ethical and Societal Concerns

  • Multiple commenters express strong discomfort and “Absolutely Do Not Want” reactions, especially around:
    • AI-only interviews, HR interactions, and training replacing human contact.
    • Call-center automation and further degrading human-facing services.
    • Photorealistic avatars worsening deepfake/identity-trust problems; preference for clearly non-human robots.
  • Others argue that harms are manageable and comparable to past disruptive tech (cars, nuclear power), urging focus on benefits and customer value rather than halting development.
  • Brief suspicion of astroturfing due to overwhelming positivity is raised; a moderator reminds participants this kind of accusation is against HN guidelines.