Realtime models
ByteDance Seed Seaweed APT2
NewReal-time interactive video research — an 8B model that emits a latent frame per forward pass at 24fps.
seaweed-apt.com · listing ≠ endorsement · re-checked 2026-09-03
- Operator
- ByteDance Seed
- Params
- 8B
- Model
- Autoregressive adversarial post-training (AAPT)
- Output
- 24fps streaming video (research, 736×416)
- Access
- Research release · no consumer API
What is ByteDance Seed Seaweed APT2?
ByteDance Seed's Seaweed APT2 is a research model for real-time interactive video: 'autoregressive adversarial post-training' turns a pretrained latent-diffusion model into a causal autoregressive generator that emits one latent frame (4 video frames) per single network forward pass, sustaining 24fps. It targets interactive use — real-time pose-driven virtual humans and camera-controlled world exploration — but as a research release it has no consumer API.
What ByteDance Seed Seaweed APT2 does
- 8B model that emits one latent frame (4 video frames) per single network forward pass (1NFE)
- “Autoregressive adversarial post-training” turns a pretrained latent diffusion model into a causal autoregressive generator
- Sustains 24fps real-time generation (research, 736×416 on a single H100, reported)
- Zero-shot extension to multi-minute nonstop output (reported)
- Aimed at interactive use: real-time pose-driven virtual humans and camera-controlled world exploration
Who is ByteDance Seed Seaweed APT2 for?
- Following the research frontier of realtime interactive video generation
- Exploring pose-driven virtual humans and explorable worlds, not just clip generation
- Comparing ByteDance’s realtime research line with its batch Seedance models
Pros and cons of ByteDance Seed Seaweed APT2
Pros
- Pushes a clear research milestone: a latent frame per forward pass at 24fps
- Long nonstop generation matters for live, continuous content
- Open research release with a public project page
Cons
- Research-only — no consumer API or product as of Sep 2026
- Speed figures are the authors’ own, on H100-class hardware
- Not a tool you can deploy today without significant engineering
What people are saying
Verbatim quotes from public coverage — collected 2026-09-03.
“Real-time, 24fps, nonstop, streaming video generation at 736x416 resolution.”
Frequently asked questions
What is Seaweed APT2?
ByteDance Seed’s 8B research model for real-time interactive video generation, using autoregressive adversarial post-training to emit a latent frame per forward pass.
Can I use Seaweed APT2?
Not as a product — it is a research release with no consumer API as of Sep 2026.
What is APT2 for?
Interactive, real-time use cases like pose-driven virtual humans and camera-controlled world exploration, not just one-shot clips.
MiniMax H3
Open-weight 33B video model (Hailuo 3.0) — 15s clips with native stereo audio; the substrate under this wave.huggingface.co
fal H3 Max Live
fal’s realtime-tuned H3 Max — five seconds of 768p in under three, faster than playback.fal.ai
Krea Realtime 14B
Open 14B autoregressive model for real-time, long-form streaming video — distilled from Wan 2.1, ~1s to first frame.krea.ai
Decart MirageLSD
“Live-stream diffusion” — restyle a live feed (camera, game, call, screen) frame-by-frame from a text prompt at near-zero latency.decart.ai