LoRa doesn't work as it should.

[JohnLoveJoy]

Just so you know, I compiled the experimental version of KoboldCPP and LoRas aren't working as they should, compared to the latest release and previous versions:

• Anything below strength 1 now results in LoRa information being overridden.

Character X/Y becomes an Asian girl.

• Even with strength at level 1, the expected appearance does not materialize.
• The model is also more likely to generate anatomical errors such as extra limbs.

[LostRuins]

@JohnLoveJoy how are you loading the lora? are you setting the LoRA as a fixed lora file + multiplier, or using the dynamic prompt based method?

@wbruna did anything change for LoRA recently?

[LostRuins]

I did a quick test that I normally do for a LoRA and it seems to work for me. More repro steps would help.

[wbruna]

@wbruna did anything change for LoRA recently?

There were some changes related to the layer streaming support (leejet/stable-diffusion.cpp@a7f2e03 ), but as far as I know that code path is inert if streaming isn't enabled.

We did get more disruptive changes before; 1.114.1 already included them, but I remember it caused issues (crashes) with the LoRA caching code. @JohnLoveJoy , does it happen in the same way right after you load Koboldcpp, versus after a few generations changing weights?

[JohnLoveJoy]

No. I configured the LoRa and the multiplier in the interface before launch, rather than using the runtime method.

But it doesn't lose as much information in close-up images of faces. Although with a fixed seed there shouldn't be any variation, I think.

[LostRuins]

The images do look different, so the LoRA is probably doing something with the same seed. But without knowing what the LoRA is or what its supposed to do, it's hard to tell.

I tried with a pixel art LoRA and it's definitely applying.

[JohnLoveJoy]

I'm using character LoRAs from NphSi. It's definitely applying LoRas, but something is different.
https://huggingface.co/spaces/nphSi/Lookalike-LoRA-Index

[LostRuins]

but something is different.

I don't understand, what is the issue? If you mean the seed feels different, that can happen for many reasons (e.g. backend improvements)... question is , is it broken/wrong?

[wbruna]

One thing that would help is checking if stable-diffusion.cpp's sd-cli has this issue too. 1.114.1 should behave the same as master-660-d2797b8; the current experimental builds have more recent versions.

But for a subtle issue like this, we really need a complete, reproducible example, including model quant, LLM, VAE, and parameters. Even platform changes could potentially interfere (e.g. new Vulkan shaders), so the more information you give us, the better.

[JohnLoveJoy]

Same seed. It has noticeable degradation in color and texture.

I'm using Vulkan.
Z-Image-Turbo Q4(Tested with Q5 as well).
Euler as a sampler.
Schedule in default.
The LLM used is as shown below.

[LostRuins]

Actually I thought the second one looks better.