Unverified50% confidenceFactExact time
传统Stable Diffusion使用CLIP-L/14作为文本编码器，嵌入维度768，最大token长度77，参数量约1.2亿
1
Sources
50%
Confidence
Long-term
Relevance
7/21/2026
First Seen
Sources
Related Claims
UnverifiedFlux 引入 T5-XXL 作为文本编码器，代价是推理时显存需求增加约8-16GB62% similarUnverified训练大语言模型/Stable Diffusion微调优先选显存≥24GB的卡（RTX 4090/A6000/A100 40GB），显存不足会直接OOM报错，无法通过降batch size解决时必须换卡62% similarUnverified运行Stable Diffusion XL基础版，NVIDIA RTX 3060（12GB显存）可流畅推理；Flux.1-dev完整版（FP16精度）需约24GB显存62% similarUnverified动态2比特量化相比FP16格式实现了约8倍的压缩比62% similarUnverified双编码器通常将语义压缩为768维向量，该数字来源于BERT-base的隐藏层大小，工业部署中常压缩到128或256维62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/584953API
curl https://kongchang.com/api/v1/knowledge/claims/584953MCP
get_claim(id=584953)