Advice on training Lora
I use AI toolkit, and want to train Z-Anime Lora - what would be the best approach to train it, so I can use the Lora on distilled 4-steps version?
There is no Z-Anime adapter like there is one for Z-Image Turbo - which works really well, so I'm not sure what would be the way to do that.
For training a LoRA, I would start from the non-distilled Base version.
If AI Toolkit can load the checkpoint directly, use:
z-anime-base-bf16.safetensors
That should be the best starting point if your goal is to make a LoRA specifically for Z-Anime.
If AI Toolkit cannot load that checkpoint properly, then train the LoRA on the original Z-Image Base model instead and test how well it transfers.
My recommendation would be to do a small test first: use maybe around 50 images, train a quick LoRA, then compare the result in ComfyUI on both Z-Image Base and Z-Anime Base. Generate a few test images with the same prompts and settings, and then decide which base gives you the better result for your LoRA idea.
After that, you can also test the LoRA on the distilled 4-step version. It should be tested carefully there, because distilled models can react differently to LoRAs than the full base model.
So in short:
Train on Z-Anime Base BF16 if possible.
If not, train on Z-Image Base.
Do a small 50-image test first before committing to a full training run.
Thank you for your recommendations!