Unverified50% confidenceFactExact time
Video real-time processing typically requires completing single-frame processing within 33 milliseconds, whereas VLM inference latency far exceeds this requirement, and a pure VLM solution may be 10 to 100 times slower than real-time.
1
Sources
50%
Confidence
Long-term
Relevance
8/30/2026
First Seen
Sources
Related Entities
Related Claims
Unverified业界通常以低于100ms的延迟作为音视频同步的体验基准77% similarUnverifiedLLMs frequently misjudge video pacing—for example, allocating 8 seconds to a shot that should be 3 seconds—which is an out-of-distribution generalization challenge.76% similarUnverifiedWith the accelerator enabled, MiniMax H3 takes approximately 2 minutes to generate 1 second of video, approximately 14 minutes for a 480P ten-second video, and approximately 20–30 minutes for a 768P ten-second video.74% similarUnverifiedVLA模型的推理延迟从数百毫秒到数秒不等，而典型机器人控制回路要求10-1000Hz的更新频率72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/826940API
curl https://kongchang.com/api/v1/knowledge/claims/826940MCP
get_claim(id=826940)