Non-official model, created by Ostris https://huggingface.co/ostris/Z-Image-De-Turbo
This is a de-distilled version of Tongyi-MAI/Z-Image-Turbo. It was fine tuned on images generated by Z-Image-Turbo in order to break down the turbo distillation. It can be used to train LoRAs on top of that should remain compatability with the base model, or it can be continued to finetune well beyond what you could do with the turbo model with the ostris/zimage_turbo_training_adapter.
Usage
This can be trained on directly without an adapter. It can also be used for inference directly. For inference, use a low CFG (2.0 - 3.0) and 20-30 steps. It seems to work well with CFG normalization as well.
My Remarks:
In my opinion, this version works better and more stable with zit-Loras, but at the same time the effect maybe is weaker.
Different samples began to play, which previously gave me a bad results, e.g. like res2m or dpmpp2m.
Should show well for Loras training and finetuning. Share you experience in posts and comments.
Description
FAQ
Comments (7)
Don't make same mistake I did and download this safetensor for training with ai-toolkit! It's not the correct format, ai-toolkit will download it when you run your training directly from the repo, don't change the name_or_path!
Hi, MetaGen. Could you describe in more detail what exactly the error is? Did your training fail to run, error message, or did it not complete successfully?
I achieved great results, and others have also reported that it worked for them.
@Mr_Einhorn how did you get it working?
the download kept stalling for me in ai-toolkit, i am currently trying to download the hf files using terminal instead. seems a safetensor is part of it but not the whole thing.
does it need a specific wf for this to work?
No, you should only use a CFG higher than 1, as this is the non-distilled version. Also, use more steps.
As a hint, you can refer to a workflow from the Gallery pictures.
I guess you can have the cake and eat it too