Runway Characters just shipped real-time talking avatars
Runway shipped Characters: a single reference image becomes a live, conversational AI avatar at 24fps. Here's what it changes for creators.
Runway shipped Characters, a real-time video agent that animates any single image into a conversational talking head at 24fps.
The pipeline turns audio into lip-synced face video with about 1.75 seconds from end-of-speech to first response frame.
For creators, it means an avatar that can hold a live conversation on your site, on Zoom, or inside a custom app, with zero fine-tuning.
A talking avatar that only moves when you stitch audio to lip-sync in post is one thing. One that talks back to your viewer in real time is another category. Runway just shipped that second thing.
Characters runs on Runway's General World Model, GWM-1. You drop in one reference image, yours or a mascot or a creature, and start a conversation. It generates the face on the fly, frame by frame, in HD at 24fps.
This post walks through what's actually new, how to wire one up for your own avatar, and where it earns its slot next to a pre-rendered HeyGen Avatar V pipeline.
What's new
Runway Characters is a real-time conversational video agent. It animates from a single reference image, with zero fine-tuning. It went live on the Runway API and the Runway web and mobile apps on May 4, 2026.
The hard numbers from the engineering write-up:
- 24 frames per second, HD output.
- 37 ms of model time per frame, under the 41 ms real-time budget.
- 1.75 seconds server-side latency from end-of-speech to first response frame.
- Built on GWM-1, generating frames autoregressively and streaming them as they render.
- Supports webcam and screen sharing, custom voice (text-to-voice or instant clone), tool calling, knowledge-base grounding, an embed widget, and live drop-in to Zoom, Google Meet, and Teams.
The pipeline trick: the diffusion transformer and the VAE decoder run on overlapping schedules, so decoding frame N-1 happens in parallel with diffusing frame N. That's what brings the per-frame cost under the real-time budget.
Receipts: full engineering breakdown on the Runway blog.
The hero image above was generated with Nano Banana 2.
How to put your avatar inside Characters
- Pick the reference image. This is the only input the model needs to lock identity. Use a clean, front-facing shot of your AI avatar at 1024×1024 or higher. No motion blur, no harsh side-lighting, no occluded features like sunglasses or hair across the face.
- Open characters.runwayml.com. Upload the reference, name the character, and pick a voice. Use the text-to-voice prompt field if you want to design a voice from a description like "warm, twenty-something woman, soft Nordic accent, calm pace", or paste a 10-second sample if you want to clone an existing voice.
- Attach a knowledge base (optional). Drop in markdown or text files of your product docs, content style guide, or persona bible. The character will speak from those instead of generic responses. For an AI creator avatar, this is where you put your tone-of-voice doc and your FAQ.
- Add tools (optional). If you want the avatar to actually do something (pull an order status, update a game state, push a captions overlay), define those as tools and wire them up through the Runway API. Results flow back into the conversation in real time.
- Embed or invite. One line of code drops the widget into your site. Or invite the character into a Zoom or Google Meet as a participant. The widget handles the real-time stream client-side.
[INLINE IMAGE, paired example A: a Nano Banana 2 reference portrait of your avatar, front-facing, soft natural light, premium creator aesthetic, the kind of image you'd hand to Characters as the identity input.]
Image prompt (Nano Banana 2): Editorial portrait of a 26-year-old woman with shoulder-length light-brown hair, sitting in a soft-lit home studio, facing the camera, neutral expression, premium creator aesthetic, soft natural light, no text overlays, no logos, identity-locked subject.
[INLINE IMAGE, paired example B: same subject, mid-laugh, three-quarter angle, same studio. Useful for testing whether Characters holds identity across expressions before you ship a long session.]
Image prompt (Nano Banana 2): Same 26-year-old woman, same studio, mid-laugh, candid expression, three-quarter angle, editorial photography, soft natural light, premium creator aesthetic, no text overlays, no logos.
When to use it
Characters is genuinely useful when your avatar needs to react to a viewer, not just speak at one. Live Q&A, tutoring, demo bots, embedded community hosts. It's not the right tool for scripted reels or polished long-form video where a pre-rendered HeyGen Avatar V pipeline still wins on cinematic stability and length. Use both — Characters for the live surface, HeyGen Avatar V or Seedance 2.0 for the recorded one.
Closing
Real-time avatars went from research demo to embeddable API in about eighteen months. If you've been waiting for the talk-back piece of your AI persona to be cheap and shippable, this is it. Build your reference image first, then plug it in — most of the work this year will be in the avatar's voice and knowledge, not the face. Start with /your-avatar and ship the live version on top.
FAQs
Can I use Runway Characters with a brand mascot or non-human design?
Yes. The system extracts style from the reference image, so cartoon, creature, and product-mascot inputs work the same as photoreal portraits.
Does it handle long sessions without the face drifting?
The autoregressive setup uses KV-cache eviction to keep temporal consistency over time. For very long sessions, expect the usual real-time caveat: re-lock identity if you notice drift by restarting the character with the same reference image.
What voice options does it support?
Text-to-voice from a prompt, or instant voice cloning from a short audio sample. Voice is assigned to the character once and stays consistent across sessions.
Can I run it inside a Zoom or Google Meet call?
Yes. Characters supports meeting integration on Zoom, Google Meet, and Teams. It can see and hear the call and respond in real time.
Is it cheaper than pre-rendering on HeyGen Avatar V or Seedance 2.0?
Different products. Characters is priced for live API minutes. Pre-rendered pipelines are priced per output clip. For live formats, Characters is cheaper. For recorded long-form, pre-rendered is still cheaper per delivered minute.