Broken Flux1

Well, actually, the ComfyUI Python environment I built is currently in an unrecoverable error state (it seems onnxruntime-gpu caused an issue with some extension), and I am in the process of reinstalling it from scratch. I plan to document the various problems and new discoveries I encounter along the way in my notes.

Therefore, I will likely be adding several updates to the article below as well.

In fact, my previous article was also written during that process.

Furthermore, along with the content of the previous article below, the content of an article I wrote even earlier is currently unusable (the nodes developed by ControlAlt that I used at the time are no longer available), so I plan to revise that as well later.

That said, the main topic this time is as stated in the title.

It is a quite rare situation... or rather, a case I have never experienced before, but the base model itself actually became corrupted.

The result is the noise shown in the image below,

The solution is simple and clear.

"Please download the model again, that is all"is how the story ends, but the point is the reasoning behind how I reached that conclusion.

In this case, I realized it because using Flux1 in Forge resulted in the same noisy images.

The fact that it only results in the same noise in Forge means it is not a ComfyUI issue. As a basic principle of logic, that is how it works, right?

The only common factor between the two is that they are "loading the same model file."
(More strictly speaking, one could say they are using the same Python libraries, including Pytorch and Transformers, but if the cause were there, it would be a much bigger problem than model corruption. If some general-purpose Python library were the cause of the noise, it would be a fatal bug in the library itself and could not be fixed immediately. However, in that case, users all over the world would be facing the same issue with Flux1, and it would be a global uproar, so that is impossible at this moment.)

I use four UIs simultaneously—A1111, Forge, reForge, and ComfyUI—and one of the reasons is that I can perform this kind of comparative verification.

In this case, if I were only using ComfyUI, there would be no basis for comparison, soit is highly likely that it would take a long time to reach the conclusion that "the model file itself is corrupted," which is a rare occurrence.