Meta ships Muse Realtime Avatar for live video AI characters

Meta AI Research introduced Muse Realtime Avatar, embodiment tech that turns Muse Realtime Voice into expressive interactive avatars — streaming 448×768 video at 25 fps with about 870 ms latency and up to 12 concurrent sessions per NVIDIA GB200 — with Meta Video Seal watermarks.

SOFTWARE desk — chat agents are growing faces that hold a character across a live call; watermarking is Meta’s receipt that the clip was machine-made, not a guarantee against misuse.

What the page says the avatar can start from. Conditioned on reference media, it brings any character into a live conversation. A photographic portrait responds through subtle expressions. A full-body illustration gestures and shifts posture as it speaks. Animals and everyday objects become expressive without losing what makes them distinctive. Frame by frame, the page says, the avatar’s appearance and mannerisms remain coherent from one conversational turn to the next. A later line says this goes beyond talking heads: any image comes to life in real time, with expressive facial, hand, and full-body movement. Reference media means the picture you hand the model. These lines are the blog’s. This desk did not upload a portrait.

How voice and face stay on the same clock. Muse Realtime Voice and Muse Realtime Avatar form a single streaming system connecting intelligence, voice, and embodiment. Muse Realtime Voice provides the conversational intelligence and produces a stream of speech tokens, which the page shortens to VQs, carrying both what is said and how it is delivered. An audio decoder turns those tokens into speech. Muse Realtime Avatar consumes the same stream to generate the matching visual performance. Sharing this token stream, the page says, keeps voice, lip motion, and expression synchronized. A VQ, in plain words, is a compressed chunk that carries the words and the delivery, not a finished video frame. These lines are the blog’s.

The model, in the page’s words. Muse Realtime Avatar is an audio-driven Diffusion Transformer. It is conditioned on the speech-token stream, the reference media, and a rolling window of recent video latents. It generates video in short causal chunks. As each chunk finishes, its newest generated latents become motion context for the next chunk. That carry-forward, the page says, keeps the avatar’s appearance and mannerisms while the computation stays bounded, so generation can continue for as long as the conversation does. A Diffusion Transformer is the model that paints each frame from noise, guided by those tokens and by the recent frames. A latent is a compressed stand-in for a frame, not the pixels you see. Causal means each chunk can use what already happened, and cannot peek at video that has not been made yet. A rolling window means only the recent stretch, not the whole call. These lines are the blog’s. This desk did not run the model.

How they made it fast enough, and which preference sentence is which. The page says live streaming has to do two things at once: make video fast enough for a real-time conversation, and stay visually consistent without errors piling up. They start with a high-quality bidirectional teacher and produce a causal student with a fixed-length KV cache, through self-forcing and distribution matching distillation. Self-forcing, the page says, lets the student train on its own generated context so it learns to resist drift as small errors add up, matching what it will see when it is actually running. A teacher, here, is the slow accurate model. A student is the fast copy. Bidirectional means the teacher can look both ways through a clip. A KV cache is a short memory of recent frames so the student does not re-read the whole call. The teacher uses 40 diffusion steps with three-way classifier-free guidance, which the page shortens to CFG. That guidance takes three model passes per step, totaling 120 model evaluations per chunk. The recipe distills the diffusion process and the effect of that guidance into an unguided two-step student. The page’s sentence is that the student reproduces in two unguided evaluations what the teacher refines over 120 guided evaluations, a 60× reduction, while closely preserving teacher quality. 60× is 120 divided by 2. The page prints 60x. A figure caption on the same page says distillation reduces inference from 120 to 2 neural function evaluations per chunk, a 60× reduction, while yielding a near-even split in overall preference. Closely preserving teacher quality is the prose. A near-even split is the caption. Do not collapse them into a win the caption did not print. These lines are Meta’s. This desk did not score the teacher against the student.

The live-call comparison, and the exception on the same page. To see how Muse Realtime Avatar did in a live conversation, Meta compared it with Runway Characters and HeyGen LiveAvatar. The page calls those two the leading commercial avatar systems. That comparative is Meta’s. Raters used each product’s native live-call experience. They held two- to three-minute conversations with each system, using matched avatar identities, then compared visual quality, synchronization, character consistency, and mannerisms, among others. The page says the chart reports overall preference, and that raters more often preferred Muse Realtime Avatar. A later sentence says raters preferred Muse Realtime Avatar overall and across every evaluated dimension. The next sentence says the mannerism comparison with Runway Characters was not statistically distinguishable from parity. Parity means a tie. Every dimension is the page’s phrase, and the Runway mannerism line is the exception the same page prints. These are Meta’s raters and Meta’s study. This desk did not sit in on a call, and the page does not print a rater count.

The speed and the capacity, as numbers with a meaning. Muse Realtime Avatar streams 448×768 portrait video at 25 frames per second. The page prints 448x768. Portrait means taller than wide. 25 frames a second is ordinary video motion. Latency is approximately 870 ms, measured from the end of a user’s turn to the first byte of the synchronized voice-and-video response. 870 milliseconds is a bit under one second. The page’s section title calls that subsecond latency. The caption on the conversation strip says the examples are delivered at interactive, subsecond latency. Subsecond is the page’s word for that 870 ms measurement. It is not a second, tighter clock. On a single session on a GB200, each generation step produces eight frames, which the page says is 320 ms of playback, in 20 ms. The page calls that an effective 2.5 ms of model time per frame. Eight frames at 25 frames a second is that 320 ms. Eight frames in 20 ms means the step finishes well before those frames have played. The page says the optimizations, plus orchestration across voice and video, increase serving capacity by 8× relative to the two-step BF16 baseline, enabling 12 concurrent real-time sessions for the video generation on a single GB200. The page prints 8x and GB200. 8× is against the already-fast two-step baseline, not against the 120-evaluation teacher. BF16 is a compact number format used to run the model. Concurrent means twelve video sessions at once on that one chip. The page names NVIDIA as the partner on model optimizations. It does not print the longer string NVIDIA GB200. The kernel, cache, and four-bit recipe behind that 8× is parked in the source note. These lines are the blog’s. This desk did not time a session.

The receipt, and who the product is for. The page says Meta enforces strict safety requirements through the experience to reduce misuse and protect people. To make generated media traceable, Muse Realtime Avatar uses Meta Video Seal to embed a durable, invisible watermark throughout the generated video, without adding latency. A watermark, here, is a hidden mark in the pixels, not a logo in the corner. Invisible means a viewer does not see it. Durable means it is meant to survive ordinary copying. Without adding latency means the mark does not slow the 870 ms path. The page says Meta will keep strengthening these protections as embodied AI evolves. All examples in the post illustrate model capability and do not all reflect avatars available in the Muse app. Muse is for users aged 18+. These lines are the blog’s. A watermark is not a guarantee against misuse. This desk did not test whether the mark survives a re-encode.

Plain English for the rest of the card. VQ means a speech token, a compressed chunk of what was said and how it sounded. A Diffusion Transformer paints frames from noise using those tokens. A latent is a compressed frame the model passes forward. Causal means the next chunk cannot see video that does not exist yet. CFG is classifier-free guidance, the teacher’s extra steering, three passes per step. 40 steps times 3 passes is 120 evaluations. The student does 2. 60× is that cut. Near-even is the caption’s preference split between student and teacher, not the Runway comparison. 448×768 is the portrait size. 25 fps is the frame rate. 870 ms is the wait after you stop talking, until the first byte of voice and video. Eight frames in 20 ms is one generation step on a GB200. 2.5 ms is that step divided by eight frames. 8× is serving capacity versus the two-step BF16 baseline. 12 is how many real-time video sessions the page says fit on one GB200. Video Seal is the invisible watermark. 18+ is the age line. 23 Sep 2026 is the page date. No hour is what this desk read. Nine conversation clips are on the page. This desk did not play them.

PRIMARY here: Meta AI Research’s 23 Sep 2026 blog, “Bringing Your Muse to Life” — Tier A PRIMARY, the lab’s own post. The page does not print an hour. The introduction of Muse Realtime Avatar, the reference-media lines, the shared speech-token stream, the Diffusion Transformer and causal chunks, the 120-to-2 distillation, the closely-preserving-teacher-quality sentence, the near-even preference caption, the Runway Characters and HeyGen LiveAvatar comparison, the overall-preference line, the every-dimension line, the Runway mannerism parity line, the 448×768 at 25 frames per second, the approximately 870 ms latency, the eight frames in 20 ms, the 2.5 ms per frame, the 8× versus the two-step BF16 baseline, the 12 sessions on one GB200, Meta Video Seal, the 18+ line, and the note that blog examples do not all match avatars in the Muse app are that page’s. NOT claimed: an hour, a price, a ship date inside the Muse app, a rater count, that the mannerism comparison with Runway was a win, that the student beat the teacher in the caption’s preference split, that 8× is a cut from 120 evaluations, that the page prints the string NVIDIA GB200, that this desk placed a call or timed a frame, that a watermark stops misuse. Distinct from the already-filed glia-grm-branch, collibra-maestro, brain-company-memory, meta-rayban-audio-glasses, meta-vr-glasses, google-gemini-38-flash-tts, wisdomai-live-apps, meta-muse-personal-ai-agent, meta-muse-0day, and amazon-blocks-meta-muse.