@0l1v1aR0551 it's not your work, and of all people on this site to try and explain the difference of BF16 to FP16, I'm not that person.

So unless you halved it yourself (and if you did, I apologize), credit the Repo at the very least.

@0l1v1aR0551 If you are using Z-Image, then you shoudln't be using FP16. Kinda defeats the purpose.

I apologize. I did not catch the notes in the releases. My bad.

@0l1v1aR0551 That's not the difference between bf16 and fp16, and chopping up somthing from fp32 is kind of the same thing. you don't need scaling for bf16 and it trains better. You're not going to create 8k images with ZImage.

Regardless, Thanks for posting the abliterated Text Encoder.

@0l1v1aR0551 train on it, run it on older machines, run it on low vram. It makes slightly better images, but it is not optimal for the model.

@0l1v1aR0551 sigh 😔.

FP16 has no range to train and is for precision only. it does not perform well at all on newer model graphics cards. Unless you are trying to make commercial grade images (which you should be using FP32 for), it's not a preferred model. I don't know if you understand tensors, but here's a video to help explain what I mean. https://www.youtube.com/watch?v=oOYUS56pcwA

It's common for people to bring other people's work to CIVITAI. In the case of image generation models, this is often permitted due to licensing issues. This case is better because the creator has arbitrarily adopted their own license and doesn't impose any restrictions. Even if the creator attempts to upload it, it may be invalid if it's already been uploaded.

The reason it wasn't uploaded to CIVITAI is because it's not worth uploading. The effort required for conversion and the internet usage required for downloading are simply a waste. If it had been converted from the FP32 version, it might have been possible to position it as a model with slightly higher mantissa accuracy, but Z-Image only releases the BF16 version. This reduces the performance of the exponent and prevents the accuracy of the mantissa from increasing. ComfyUI and other tools assume BF16 processing, further reducing efficiency.

The FP16 version would be a boon for those using image generation AI with GeForce 2K-generation GPUs and CPUs. Unfortunately, there aren't any such people.

Some argue that FP32 provides better accuracy, but converting from BF16 to FP32 is completely pointless. Incidentally, I don't really understand why some people are happy about AIO versions. If you switch models, you have to reload the text encoder and VAE, right? Even if those models are pointless, they're not prohibited from being uploaded.

@aueki4g467 so this wasn't even a halved model?

@lonecatone23 fp16 is not halved from bf16, some times it is halved from fp32 but there is no fp32 model yet. So this model might be converted from bf16 version, this meas who created this needs to consider how to rebuild some data from bf16 to fp16 losslessly.

@Garland I'm aware how it works. That's why I'm surprised this sven exists. There is an FP32 version out there as well, which makes me wonder how they created that.

Sounds like trying to put mashed potatoes back together

Oh hey, that's my HF!

So, this model was created when the "FP32" version of Z Image was found deleted from the Tongyi-MAI repository that had 3 safetensors files instead of the 2 file version we all use as BF16. I only made this for those who could run the FP32 that was being posted all over the place, and they needed something older than BF16 for the same reason.

Considering that all the models we're talking about here are Apache 2.0, and 0l1v1aR0551 even linked to my HF, it's all good.

I didn't upload it myself to Civitai because my upload is only 20Mbps. It was a huge lift just to get it onto HF.

I never went deeper into testing it after making some XY plots. It did teach me that PyTorch auto-converts both this and the full FP32 into BF16 when running in Comfy on my GPU.

No official word has come to light why Tongyi-MAI deleted that commit and replaced it with a smaller version later.

Thanks 0l1v1aR0551 for reposting, but maybe this entire thread could have been avoided by copying the readme, so people would know this is from commit 1ac663b and not the later commit we all usually use. This was created from the early version which can still be found here: https://huggingface.co/Tongyi-MAI/Z-Image/commit/1ac663bf5fcfa5b4752f9e1315cbb5a4541ed7b8

@0l1v1aR0551 I updated the HF readme to link back to here, so hopefully the confusion can be cleared in at least one direction.

Crazy to see everyone here is thinking I made it from the BF16, but it was converted from the 24.6GB version that was mysteriously deleted and found in the old commits. So I hope the air can be cleared that this is from completely different model layers.

@070809 That clarifies a lot. There was no link to the model before. Now we know. Thanks!

@0l1v1aR0551 To this day there is only one pic and no notes in the comments, where 99% of legitimate posters put their links. I should not have to go and dig through a highly suspicious post to figure out where it came from. Several other people also asked, so I was not alone with this issue. Further, ther was no explaination where the FP16 version come from and only the BF16 version was officially posted. It looked suspicious.

I'm glad it has been clarified and I already apologized for the confusion. Next time be more open about your sources.

@0l1v1aR0551 dude, give it up. You have one pic and no description and have been called out numerous times by several different people. There are numerous scammers out there. How about posting it where it belongs on the main page? Not hard.

Also, at the end of the day it still wasn't your work.