January 27, 2026
FaceTime with bots? HN’s got thoughts
Show HN: LemonSlice – Give your voice agents a face
HN tests talking avatars: wow, weird, and work-ready
TLDR: LemonSlice launched real-time video avatars so you can FaceTime with a chatbot, opening a playground and API. HN’s comments mix excitement (“I loved it!!”) with practical asks—plugging in OpenAI voice, emotion controls, and training use cases—debating whether friendly faces can beat plain text and the lingering uncanny vibe.
LemonSlice wants your chatbot to stop hiding behind text and actually show up on video. The founders dropped “Lemon Slice 2,” a model that streams photoreal faces at 20 frames per second on one graphics card, and opened an HN playground. They admit the uncanny valley isn’t conquered yet—cue the crowd’s split reaction. Early testers like zvonimirs are already turning assistants into “coworkers” that record Loom-style clips and shouting “I loved it!!” Meanwhile the hype squad chimed in with “This looks super awesome!” and dared friends to chat with the mood-boosting alien.
Then the practical questions hit: buddycorp asked if they can “plug in my own OpenAI realtime voice agents,” while dreamdeadline pressed for fine-grained control over movement, expressions, and emotion. Founder sid-the-kid revealed surprise traction in roleplay training—think nurses triaging AI patients and sales reps practicing leads. Cue jokes about reviving Max Headroom and office avatars that never call in sick. With a demo, docs, and a free month coupon floating in the playground, the vibe is: ship fast, make it less creepy, and let people build. If faces beat text, HN’s ready to talk to bots like they talk to coworkers.
Key Points
- •LemonSlice released Lemon Slice 2, a 20B-parameter diffusion transformer for real-time avatar video generation.
- •The model produces infinite-length video at 20fps on a single GPU and supports photorealistic, animal, and stylized cartoon avatars.
- •Real-time streaming was enabled by making the model causal and optimizing memory with sliding window attention.
- •Speed and quality improvements included distilling denoising steps with GAN-based methods and inference optimizations such as real RoPE, kernel fusion, and a rolling KV cache.
- •The API is open with usage-based pricing ($0.12–$0.20 per minute), a LiveKit integration, demos, and an HN playground with a coupon for a free first Pro month.