이미지: X — 미디어·생성AI
Summary
- LTX.io revealed that its real-time conversational avatar 'Reactor' (@reactorworld) streams video frame-by-frame on the fly using LTX technology
- Rather than playing back pre-made video, the key point is that the screen doesn't freeze while a response is being generated
- LTX unveiled its open-weight video model LTX-2.5 on August 11, demonstrating the ability to generate a 10-second clip in 6.8 seconds on RTX GPUs
- 발표 계정
- LTX.io (X, 2026-08-14)
- 적용 사례
- Reactor (@reactorworld)
- 핵심 기술
- 프레임 단위 실시간 스트리밍, 사전 렌더링 없음
- 용도
- 학습(러닝) 경험 구축
- 배경 모델
- LTX-2.5, 2026년 8월 11일 공개
- 온프레미스 생성 속도
- 10초 클립을 6.8초에 생성
- API 생성 속도
- 10초 클립에 23.7초
An avatar that freezes isn't real
It's a familiar scene: a conversational AI avatar's screen briefly freezes while it's coming up with an answer. That gap — where the avatar pauses for a few seconds after being spoken to before moving again — is exactly the "uncanny valley," the zone where something resembles a human closely enough to feel unsettling rather than convincing. Video generation startup LTX announced via X on August 14 that it has unveiled a real-time conversational avatar that eliminates this gap. The application cited was the conversational avatar service 'Reactor' (@reactorworld).
What frame-by-frame real-time streaming means
LTX explained, "LTX streams video frame by frame, not by pre-rendering and then playing it back." Typical video generation AI computes an entire multi-second clip before showing it to the user — and that computation time is exactly what users experience as a "freeze." What LTX emphasized is that it has eliminated this compute-then-playback structure altogether. Because the screen continues to stream frame by frame even while a response is being generated, the avatar appears to be carrying on an actual real-time conversation. Reactor was mentioned as using this speed to build learning experiences, but the specific form of the service or its target age group and subjects were not disclosed in this announcement.
Background: LTX-2.5's speed as the foundation
As METAL LAB reported on August 12, LTX unveiled its open-weight video generation model LTX-2.5 on August 11, lowering VRAM requirements to enable local operation on consumer-grade NVIDIA RTX graphics cards and DGX Spark. At the time, on-premises benchmarks (using 2x GB200) showed it generating a 10-second clip in 6.8 seconds, more than three times faster than the 23.7 seconds needed to produce the same clip via API. The real-time avatar feature unveiled this time appears to be a real-world application of that speed advantage. Streaming frames instantly, rather than computing an entire clip before delivering it, is far more advantageous for conversational services.
Comparing avatar response methods
| Method | Processing flow | Perceived latency |
|---|---|---|
| Conventional method (pre-rendering) | Compute entire clip → play back | 70 |
| LTX frame streaming | Compute and play back frame by frame simultaneously | 15 |
The bars in the table represent a conceptual comparison of the relative perceived latency emphasized in this announcement, not separately published benchmark figures.
What Reactor is trying to do
Reactor was mentioned as using LTX's technology to build learning experiences. Avatars with uninterrupted conversation flow make a particularly noticeable difference in educational services where real-time interaction matters, such as language learning or tutoring — since it can prevent learners' attention from wandering while waiting for a response. However, whether Reactor's service has launched, how to access it, and pricing details were not addressed in this tweet.
So what changes
This is a signal that video generation AI is shifting from a tool that "makes everything first and shows it later" to one that "keeps making it while showing it." Until now, conversational avatars often generated a text or voice response first and then layered video on top to match it. But if frames themselves can be streamed in real time as LTX does, the applications extend beyond avatars to live broadcasting, remote consultations, and video-based education — any service where latency itself determines quality. Given that LTX-2.5 has been released as open weights, cases like Reactor's could lead to further applications by other developers as well.
