how can I select weight type in Krita

[one903045]

how can I select weight type in krita like comfyui,Krita default generator use default BF16, hight VRAM for me. I don't know about the FP8 weight type OF BF model AND Native FP8 model Whether they have the same quality. maybe i just use Customize workflow or download FP8 model to use.

[CommunicationDesignation]

I believe this is all done automatically. Only in comfyui would you need to manually change the weight type to fit the model.

This is the only place you can chose from automatic or manual, but weight type is not a consideration. Id guess that's something that could easily be made automatic.

Is that bf16 model not working for you?

[CommunicationDesignation]

If you look at the required model list you'll see the only bf16 model is SDXL.

Also, this link below might be helpful to get it working. But again I don't see anything about bf16 qwen models. bf16 are huge in file size, qwen being the newest and one of the largest in memory requirements, Krita-ai-diffusion might not be set up for such large requirements. Its just a guess though.
Krita-AI handbook:
https://docs.interstice.cloud/edit-models/#supported-edit-models

[one903045]

I believe this is all done automatically. Only in comfyui would you need to manually change the weight type to fit the model.我相信这一切都是自动完成的。只有在 Comfyui 中才需要手动更改体型以适应模型。

This is the only place you can chose from automatic or manual, but weight type is not a consideration. Id guess that's something that could easily be made automatic.这里是唯一可以选择自动或手动模式的地方，但重量类型不在考虑范围内。我猜想这部分很容易就能实现自动化。

Is that bf16 model not working for you?你那款 BF16 模型是不是用不了？

I can use bf16 working normally, I mean is that I hope to convert to FP8 weights to reduce the use of VRAM like comfyui.Otherwise, I will need to keep both models BF16 and FP8.

[Acly]

There is no way to select weight type explicitly. If you want BF16, use a BF16 model, if you want FP8, use a FP8 model. Usually you use whatever fits best, it doesn't make sense to store a BF16 model on disk if you then load it as FP8.