flux1-krea-dev_float8_e4m3fn_scaled_learned_stochastic.safetensors
Back

flux1-krea-dev_float8_e4m3fn_scaled_learned_stochastic.safetensors
0 0 23

Realistic

flux1-krea-dev_float8_e4m3fn_scaled_learned_stochastic.safetensors

flux1-krea-dev_float8_e4m3fn_scaled_learned_stochastic.safetensors

(解释来源于Deekseek,仅供参考)


scaled_learned_stochastic:

scaled:在量化前后会做比例缩放(scale factor),减少溢出和精度损失。

learned:量化缩放参数或分桶范围是在训练中学得的,而不是静态计算。

stochastic:量化时加入随机性(stochastic rounding 或随机抖动),在推理阶段有助于减轻量化误差的偏差模式。

特点:这种方法通常在生成类任务中可以保持更高的多样性,有时输出会更“活”,但稳定性和可重复性略低于 SVD 方案。


核心差异总结:


文件名 方法 重点优化 输出特点
learned_svd 奇异值分解+可学习参数 权重低秩近似 精度高、稳定性好
scaled_learned_stochastic 缩放+可学习随机量化 抗量化噪声 & 多样性 多样性高、可能稍微不稳定

如果你想推理一致性高、结果可控 → 用 learned_svd
如果你想生成更有创意、接受一定不稳定性 → 用 scaled_learned_stochastic
This model is sourced from an external transfer (transfer address: https://huggingface.co/Clybius/FLUX.1-Krea-dev-scaled-fp8 ),if the original author has objections to this transfer, you can click,
Appeal
We will, within 24 hours, edit, delete, or transfer the model to the original author according to the original author's request

AIGCTV

AIGCTV

Realistic

Model Information

Original author:
Clybius
Model Type:
Unet
Basic Model:
Flux.1-Krea
Resource Name:
models/unet/flux1-krea-dev_float8_e4m3fn_scaled_learned_stochastic.safetensors
MD5:
193318938152fa75ccd599af3f87b96d

flux1-krea-dev_float8_e4m3fn_scaled_learned_stochastic.safetensors

(解释来源于Deekseek,仅供参考)


scaled_learned_stochastic:

scaled:在量化前后会做比例缩放(scale factor),减少溢出和精度损失。

learned:量化缩放参数或分桶范围是在训练中学得的,而不是静态计算。

stochastic:量化时加入随机性(stochastic rounding 或随机抖动),在推理阶段有助于减轻量化误差的偏差模式。

特点:这种方法通常在生成类任务中可以保持更高的多样性,有时输出会更“活”,但稳定性和可重复性略低于 SVD 方案。


核心差异总结:


文件名 方法 重点优化 输出特点
learned_svd 奇异值分解+可学习参数 权重低秩近似 精度高、稳定性好
scaled_learned_stochastic 缩放+可学习随机量化 抗量化噪声 & 多样性 多样性高、可能稍微不稳定

如果你想推理一致性高、结果可控 → 用 learned_svd
如果你想生成更有创意、接受一定不稳定性 → 用 scaled_learned_stochastic
This model is sourced from an external transfer (transfer address: https://huggingface.co/Clybius/FLUX.1-Krea-dev-scaled-fp8 ),if the original author has objections to this transfer, you can click,
Appeal
We will, within 24 hours, edit, delete, or transfer the model to the original author according to the original author's request