mflux macos
i have this error using the Loras,
✅ Applied to 0 layers (0/630 keys matched)
⚠️ 630 unmatched keys in LoRA file:
- transformer.layers.0.attention.to_k.alpha
- transformer.layers.0.attention.to_k.lora_down.weight
- transformer.layers.0.attention.to_k.lora_up.weight
- transformer.layers.0.attention.to_out.0.alpha
- transformer.layers.0.attention.to_out.0.lora_down.weight
... and 625 more
✅ All LoRA weights applied successfully.
I use z-image turbo mlx 4 bit
also the image generated dont respect the lora,
maybe I have to use Z-image base ?
What app do you use to generate image? It works with Turbo but i never tested anything else than FP8 and Q versions. Afaik 4bit is Blackwell only. It will work in Comfy, anything else uncommon is not tested by myself.
Invoke works but you have to manually config the Loras because Comfy fucked up the file format because they wanted to be the first to support ZI.
thanks for the reply , I use Mflux , its a mlx metal for macOS : https://github.com/filipstrand/mflux , anyway my question is, your lora can works with turbo? or just zimage base? Because looking at your lora I see : "training_method": "LORA", "model_type": "Z_IMAGE", btw thanks.
They are all trained on Base and work with Turbo and Base.
If you do not get it to work you should open an issue there telling them that Onetrainer Loras do not work because they are saved in standard format and not longer in the comfy specific hack.
Sorry that i can not help any further...
ok , thanks for all.