feng wang

Hi Vitaliy, Sorry for my late reply. 1. You can simulate the dataset `simulated_dataset/simulated_stem_images_128x128_16bits` using the script https://github.com/fengwang/Noise2Atom/raw/master/code/simulate_physical_model.py. Just remember to update the path to the folder for the simulated...

@vkinakh Please find the first 100 frames of it from this place: https://github.com/fengwang/Noise2Atom/releases/download/experimental/2019_03_26-15.35abcdf.2048px.8.8pm.tif.too_large_0_100.tif

> This URL is no longer working. > > https://drive.google.com/file/d/19PnyLngHPV55LBYHJeAi6wXTDmzTeXKv/view?usp=sharing > > Could you please provide it again? Have you tried the data from https://github.com/fengwang/Noise2Atom/releases/download/experimental/2019_03_26-15.35abcdf.2048px.8.8pm.tif.too_large_0_100.tif

It is expected as this project will produce images containing 2048x2048x3 pixels in 20 steps by default.

https://github.com/fengwang/Stable-Diffusion-NCNN/blob/e3ba9d807c0c40197c8b13be2d94c690b52f7045/stable_diffusion.hpp#L784 字典啊

似乎可行。 但是这个模型是 dynamic quantization 的，也就是只有权重被 quantize 到了 int8，模型中间层的输入输出还是 fp32。 我稍后尝试下，看运算速度能提升多少，还有输出是不是会有瑕疵。

Do you have pkgconf and ncnn installed? What is the output of `pkg-config --cflags --libs ncnn`?

ncnn 不自带的话只能自己仿照这个（https://github.com/atanmarko/ncnn-with-cuda）造个轮子了。但是我不是很看好，因为工程量太大了。

Please note that I encountered this or same problem with many versions, from 0.58.6 to latest

Hi @zadam, many thanks for looking into this issue. I created the docker container in a server in my Lan, and visit it remotely from my laptop using a zerotier...