nphSi

/

Z-Image-Lora

like 26

Haven't tried them yet (I know, this isn't really helpful) but as soon as the finetunes (the good ones) are out, I will check base out. Until then I'm very sufficient with the turbo loras. Again, thank you very much for your work!

Its not about checking Base, its about checking the Loras i trained on base with normal Z-Image Turbo. Nothing changes for you. Filenames change from ZIdT to ZI. ZIdT Loras are trained on deTurbo, ZI on new base.
Base is meant for training/finetuning, although you can generate with it but 50 Steps and quality is very "basic".
But you are right with Finetunes, they will be based on "Base" and hopefully Turbo Verisons of them will come, or some kind of DMD Lora.

As for now i will train on "Base" only since i believe i found settings that surpasses deTurbo quality. Therefore i need the feedback.

BTW, please use the 'official unofficial' FP32 weights of Base to train on. You can find them: https://huggingface.co/Hellrunner/z_image_fp32
(recovered from the originals uploaded to Tongyi repo and then deleted, but in commits, replaced by [lesser] 16bit official version.

FP32 trains better. Try it and compare.

Can you please share some output samples of fp16 vs fb32 trained lora on same dataset?

I have more or less an output LIGHT when I compare your sample images: As far as i understood, you trained Jessica Alba and Jennifer Aniston again on Z-Image Base. That's why there are two versions each currently. But when I compare the sample (!) images, the new Z-Image Base Loras look a bit off. Jennifer is a bit less Jennifer and Jessica is bit less Jessica. It reminds me a bit on Flux Klein where the likeness is 85 % while ZIT has 95 %. That's just my "feedback". According to Malcom, there are indeed problems when it comes to Z-Image Base Loras: https://www.reddit.com/r/StableDiffusion/comments/1qv6o5q/z_image_vs_z_image_turbo_lora_situation_update/

This is not a criticism, quite the contrary, just a comment. Keep up the good work!

Yes its currently a big process of finding the right settings to train on base. Doing multiple Loras every day with different settings and combinations while being limited to 8GB Vram. I just post the "best" results to keep track of them, they are not meant to be final. I will re-train everything anyway when working settings are found. Its a bit demotivating atm. Same with F2k...

For me, there is currently no competitor regarding ZIT when it comes to Loras and likeness not even within their own camp (Z-Image Base).

As long as its not clear how to train on base i am going to train the few newer Loras from the last days on de-Turbo also since they just work better atm. I really hope it gets sorted out soon.

Edit: Settings found, training on base...

Can you please share some output samples of fp16 vs fb32 trained lora on same dataset?

apologies for not providing samples yet, I asked a few of the people who have been doing both and gotten no examples back yet, but consensus is still that training on FP32 is better. overall. I've even reached out to Tongyi and asked they officially release the fp32 weights.

Hi, I think I mentioned before that your ZiBase trained loras are performing MUCH better on Turbo than the original Turbo-trained loras. Well done!
I have only tried maybe 5-6. One exception I spotted comes out cartoony, smooth, often with glitched anatomy. Not sure if you tried different settings or if there was a dataset issue, etc. But wanted to let you know that one was producing odd results compared to others.

Yes i noticed that too, had to replace some images in the dataset. Looks like the model thinks its a cartoon due to her big eyes 😃 Her whole proportions are somehow "special"...
New version will be up in a few hours.
Thanks for feedback and have a nice evening!

Edit: retrained and up.

Thank you!