【光影扩充】IC-Light整合包

Cxsmo

does this support directml or no

Cxsmo

Is there a way for you to add the ability to have a full folder of backgrounds and then a dfl unpacked faceset and then output the dfl faceset with a random selection of the backgrounds so basically way better random color

lispmox

Cxsmo 发表于 2024-10-8 16:59
Is there a way for you to add the ability to have a full folder of backgrounds and then a dfl unpack ...

Modify the bat script to add the --max-bg parameter. When the number of images in the backgrounds folder exceeds max-bg, each aligned face will randomly select max-bg background images for enhancement. For example, --max-bg 8 means that each aligned face will randomly select 8 background images for enhancement.

lispmox

Cxsmo 发表于 2024-10-8 16:56
does this support directml or no

目前看没有技术难度，目前主要问题是处理速度并不理想。directml版本的pytorch缺少memory efficient attention的实现，在运行时会OOM。我也参考了comfyui采纳的--sub-quad-attention，但在我的卡上运行很慢，并且生成的图像出现了一些破损。所以我目前不准备发布这个半成品的directml版本。如果大家有更好的可以替代torch.nn.functional.scaled_dot_product的directml方案，请告知我，我会尽快发布可用的版本。

Currently, there is no technical difficulty, but the main issue is that the processing speed is not ideal. The DirectML version of PyTorch lacks the implementation of memory-efficient attention, which results in out-of-memory (OOM) errors during runtime. I also referenced the --sub-quad-attention adopted by ComfyUI, but it runs very slowly on my GPU, and the generated images have some artifacts. Therefore, I am not planning to release this unfinished DirectML version for now. If anyone has a better DirectML solution to replace torch.nn.functional.scaled_dot_product, please let me know, and I will release a usable version as soon as possible.

lispmox

丸丸 发表于 2024-10-2 00:40
请问是一键包吗，需不需要装其他东西

打包好了，解压就可以用

Cxsmo

lispmox 发表于 2024-10-9 15:57
目前看没有技术难度，目前主要问题是处理速度并不理想。directml版本的pytorch缺少memory efficient atte ...

what about zluda

lispmox

Cxsmo 发表于 2024-10-10 00:56
what about zluda

I took a quick look at ZLUDA, and it doesn't seem to be ready for plug-and-play use at the moment. I also don't have an AMD card for testing. Perhaps I should consider using solutions like the ComfyUI workflow.

forumphp

rem ========== PROXY ENV ==========
set HTTP_PROXY=http://127.0.0.1:7890
set HTTPS_PROXY=%HTTP_PROXY%

丸丸

楼主最近除了flux的模型，要不要试试呢

孤独朋克男

买完余额：4 希望能用