Unverified50% confidenceFactExact time

Krea团队在模型发布时同步提供了经过专项优化的FP8权重，而非简单对BF16权重进行事后量化

1

Sources

50%

Confidence

Medium-term (~90 days)

Relevance

7/10/2026

First Seen

Valid until: 10/8/2026

Sources

Krea2实测：FP8与BF16画质差距还存在吗？

redditr/StableDiffusion7/9/2026

Related Claims

Unverified模型量化是将模型权重从FP16压缩至INT4/INT8以节省显存71% similar Unverified生产级实现如fplll通过MPFR库实现任意精度浮点运算，用于精度回退机制64% similar UnverifiedNVFP4引入每组缩放因子机制，通常以128个权重为一组共享一个FP8缩放因子64% similar Unverified全量微调会更新模型所有权重，效果最佳但对算力要求极高64% similar UnverifiedDiffSynth-Studio支持模型分块加载、半精度（FP16）推理以及显存与内存之间的智能调度等优化技术63% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/465378

API

curl https://kongchang.com/api/v1/knowledge/claims/465378

MCP

get_claim(id=465378)