Compatibility:
Sorry for the extremely long post.
As anyone familiar with the Stable Diffusion community can somewhat understand, image generation in HF once rose to prominence and then fell into disuse, in part because the A1111 WebUI became the de facto standard.
The single safetensors file became the norm, making it somewhat incompatible with Diffusers.
The rise of Civitai was probably not the only reason.
(I wasn't yet playing with AI at that time, so I can only speculate...)
As far as common architecture models are concerned, it is possible to convert them to each other by giving up metadata and some information loss, and conversion scripts are also available. In fact, there are probably people who do so.
The internal physical differences are at best just the names of the keys, not the substance...
But it's a pain in the ass and enough of a problem to abandon the site.
As you know, FLUX.1 is currently very popular and is always relatively high on the trend list.
So what is the format used for the distribution and exchange of this model, and it is not even a stand-alone five-body safetensors. Even within HF, where there is plenty of storage, but even more so on Civitai and other sites where capacity is tight.
Specifically, it could be a torch.float8_e4m3fn, or it could be the distribution of only Unets (Transfomers) quantized by NF4 in bitsandbytes, and so on.
But anyway, there is rarely anyone distributing in torch.bfloat16 or torch.float32, which are supported by HF's Serverless Inference API.
Not everyone has the bandwidth to upload 30GB at a time.
We can use both NF4 and torch.float8_e4m3fn from our Zero GPU space, but that's all.
We even have to write cumbersome code to use NF4. (I'll give you an example below, but it's tedious, right?)
Even though it was Mr. lllyasviel and sayakpaul at HF who started NF4 quantization.
https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/981
What I have described at length is a preamble for those who do not know the circumstances of image generation.
If this continues, it will mean that HF will be even further removed from the de facto standard in the image generation community, which is sad.
As long as the source is open, it is open, but that is not the case.
Seamless integration into the ecosystem is what is most important.
Please be the HF that explores what is possible to achieve this. The key is compatibility.
Unless you are Apple or a 16-bit era computer company...
P.S.
Today, the situation has become even more chaotic with the announcement that the next major version of the Pony model, the current de facto standard within SDXL, will support AuraFlow over FLUX.1.
In any case, we can expect more diversification of exchange methods.
Towards Pony Diffusion V7, going with the flow.
https://civitai.com/articles/6309/towards-pony-diffusion-v7-going-with-the-flow
How to use NF4 quantized FLUX.1 from Diffusers in Zero GPU space:
https://huggingface.co./spaces/nyanko7/flux1-dev-nf4/blob/main/app.py
https://huggingface.co./spaces/nyanko7/flux1-dev-nf4