MIN_SNR_GAMMA not implemented for flow matching models(ZIB-lora-Traing)

[CGPG-Wang]

Has anyone run into this error before?

I’m trying to train a LoRA using Z-Image-Base as the base model. I enabled the following settings:

"loss_weight_fn": "MIN_SNR_GAMMA"
"loss_weight_strength": 5.0

But I get the following error:

raise NotImplementedError(f"Loss weight function {config.loss_weight_fn} not implemented for flow matching models")
NotImplementedError: Loss weight function MIN_SNR_GAMMA not implemented for flow matching models

It seems that MIN_SNR_GAMMA might not be supported for flow matching models. Has anyone encountered this before or knows how to fix it?

json file:
0304-Lia.json

error screenshot

[maedtb]

MIN_SNR_GAMMA is not currently supported for flow models (flux, zimage, etc). OT currently supports CONSTANT or SIGMA for flow models.

[CGPG-Wang]

Thanks a lot for the quick reply!!

Between CONSTANT and SIGMA, which one usually works better for ZIB LoRA training in terms of loss convergence?

I previously spent quite a long time trying to train with ai-tooltik, but the loss just wouldn’t converge, which is why I started experimenting with different settings.