Can it be changed to support manual selection of custom FLUX models?

[lzhfdwu007]

FLUX.1-dev is really too big and inflexible. Most people have already downloaded the FLUX related models in independent file formats. There is no need to download the complete folder of FLUX.1-dev from huggingface again (crying). This folder is really too big...

[asutermo]

Are you on the newest version? I probably need to update docs so you won't have to download again.

There's some work to do to improve FluxFillPipelineNode. That said, one of the largest files is the thing that makes TryOff work, the FluxTransformer2DModel which is > 20 GB, and does not come with any other download of Flux.

I should have some time tomorrow to make improvements, and give more choice with regards to vae, text encoders and tokenizers. I could allow one to choose the transformer, but that will result in 'tryoff/tryon' not working.

[lzhfdwu007]

I just installed it on the 20th, so it should be the latest version, right?
models--xiaozaa--cat-tryoff-flux has 22G after downloading, but this is not the key point. The key point is that the FLUX.1-dev I pulled from huggingface has a size of more than 100G...
Of course, it would be better if models--xiaozaa--cat-tryoff-flux is not in the root directory of models.. There are too many folders in the root directory of models..

[asutermo]

Got it! Well I'm all for trying to make things easier and less bloated. And agreed, I personally have had to download flux 1 like 7 times for various nodes.

I'm working on another branch and trying to possibly support Comfy's model as opposed to diffusers.

Some patience requested (as I have to find spare time to work on this and need my desktop).

[lzhfdwu007]

Thank you for trying to reduce the hardware requirements~~~ Thank you for listening~~ THX

[304523475]

你好，我也是已经下载了独立文件格式的 FLUX 相关模型，什么时候可以支持用正常unet节点或者checkpoingt节点加载串联模型吗

[Amazon90]

nftblackmagic/catvton-flux#26

[asutermo]

Hey folks, I recently started a new job, so my time to investigate this has been minimal. I had a few branches trying some things out, but never was quite happy with it.

Comfy's way of loading models and HF's way of loading models are disparate, and while there is some code, it takes time for me to untangle, and by the time I can get back to it, I've forgotten.

I haven't forgotten, but if someone else has a change to make it work, I'd be happy to merge it.