Same setting, but the output is totally different image. fixed seed, same ksampler

[n714]

[huchenlei]

I think is the limitation of https://github.com/layerdiffusion/LayerDiffusion. Layer diffusion is essentially adding a LoRA to conv/attn layers of the model. It is not surprising that you get different images with fixed seed but with/without LoRA applied.

[n714]

I also tested on forge version, in same condition. it does show the same image out in transparent , do you meant it's the comfyui issue?

[huchenlei]

SD Forge

Without layer diffusion:

With layer diffusion:

Conclusion

I do not think is issue is specific to the ComfyUI implementation.

[n714]

yes, there is a little bit different