Unverified50% confidenceFactExact time
去噪塔借鉴图像扩散DiT架构的adaLN时间步机制，用于编码掩码比例，仅增加约150万参数
1
Sources
50%
Confidence
Long-term
Relevance
7/20/2026
First Seen
Sources
Related Claims
Unverified近似最近邻ANN算法如HNSW和IVF通过牺牲微小精度换取数量级速度提升66% similarUnverified自适应层归一化（AdaLN）由DiT架构推广，将条件信号通过MLP映射为缩放和偏移参数，相比交叉注意力计算开销更低66% similarUnverified在Prefill阶段三类模型性能差异在噪声范围内，Decode阶段打补丁后反而快约1.4%66% similarUnverifiedINT8量化、通道剪枝等模型压缩技术可在精度损失低于1% mAP的前提下将推理速度提升2-4倍65% similarUnverified将AI推理能力部署到边缘节点可以将延迟从数百毫秒降低到个位数毫秒65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/565263API
curl https://kongchang.com/api/v1/knowledge/claims/565263MCP
get_claim(id=565263)