Is Distill LoRA(For Z-image base) still not well supported at the moment?

[murasakii0118]

Currently, there is a PAI/Z-Image-Fun-Lora-Distill on HF and MS, a Distill LoRA suitable for the Z-image series models. However, when I use this LoRA in the normal way of loading LoRA, it seems to not work properly and outputs the following content.
[WARN ] lora.hpp:815 - unused lora tensor |lora.model.diffusion_model..context.refiner.1.attention.to_q.weight.lora_down|
(bf16): shape(3840, 128, 1, 1)
[WARN ] lora.hpp:815 - unused lora tensor |lora.model.diffusion_model..context.refiner.0.attention.to_k.weight.alpha|
(bf16): shape(1, 1, 1, 1)
[WARN ] lora.hpp:815 - unused lora tensor |lora.model.diffusion_model..layers.13.feed.forward.w1.weight.lora_up|
(bf16): shape(128, 10240, 1, 1)
[WARN ] lora.hpp:815 - unused lora tensor |lora.model.diffusion_model..layers.18.feed.forward.w2.weight.alpha|
(bf16): shape(1, 1, 1, 1)
[WARN ] lora.hpp:815 - unused lora tensor |lora.model.diffusion_model..layers.18.feed.forward.w2.weight.lora_up|
(bf16): shape(128, 3840, 1, 1)
.....and a long string of 'unused lora tensor' warnings, and Z-Image-Fun-Lora-Distill has no effect at all

The generation parameter template I use is
.\sd-cli.exe --diffusion-model z-image-Q8_0.gguf --llm Qwen3-4B-Instruct-2507-Q8_0.gguf --vae ae.safetensors --cfg-scale 1.0 --diffusion-fa --steps 4 --offload-to-cpu -W 576 -H 1280 --lora-model-dir "." -p "lora:Z-Image-Fun-Lora-Distill-4-Steps-2602:0.7your prompt" -o opt.png

Is Distill LoRA currently not supported, and will there be support plans in the future?
Or perhaps I'm using it incorrectly. What's the proper way to use it?

[stduhpf]

lora.model.diffusion_model..layers

the double dots are strange. Looking at the file on Huggingface, it seems they are using underscore notation and for some reason there are double underscores, that's weird. I think it would not be too hard to fix though (simply replacing double dots with simple dots should suffice I think).

But you could always use the "ComfyUI" versions, that I believe should already be compatible with sd.cpp:
https://huggingface.co/alibaba-pai/Z-Image-Fun-Lora-Distill/blob/main/Z-Image-Fun-Lora-Distill-4-Steps-2602-ComfyUI.safetensors
https://huggingface.co/alibaba-pai/Z-Image-Fun-Lora-Distill/blob/main/Z-Image-Fun-Lora-Distill-4-Steps-2603-ComfyUI.safetensors

[murasakii0118]

You're right, it worked perfectly after switching to the ComfyUI version. Thank you.