search

rgthree / rgthree-comfy

Making ComfyUI more comfortable!
MIT License
1.22k stars 86 forks source link

[Potential Bug?] Power LoRA Loader not triggering and different output #367

Open gonzalu opened 1 week ago

gonzalu commented 1 week ago

Hello,

Some users have confirmed as I have also confirmed that the Power LoRA loader is not functioning exactly as the Comfy Core LoRA loader nodes are. The output is always different and in some cases, the LoRAs have no effect on the resulting image whatsoever.

Is this expected? Is Power LoRA loader using the same Comfy Core nodes code or functionality or is it a custom LoRA loading algorithm?

Thanks for looking.

gonzalu commented 1 week ago

Comfy Core Loader image

Power LoRA Loader image

rgthree commented 1 week ago

Power Lora Loader should be the same. ComfyUI can't find a workflow in these images, do you have an example workflow?

gonzalu commented 1 week ago

MG 111a Core LoRA Loader SD3.5M Core Nodes LLM 2x fp8 Smerge 2x Upscale Conditioning Switcher McBoaty Upscale Florence2 Inspire Prompts from File AI Random LoRA Flux.1 Dev workflow.json

gonzalu commented 1 week ago

Try this one...

I may have found that the problem lies in not refreshing often. I seem to get the differences mainly when adding a LoRA and then refreshing from UI instead of the whole browser. Not sure but if I reload the entire browser, it seems to be much closer together. Not identical but also not that far off.

Thank you...

P.S. I [cannot] go back to using independent loader nodes LOL. This power loader has spoiled me and makes life soooo easy especially when mixing LoRAs :)

rgthree commented 1 week ago

Could you create an example of the issue using only a small set of nodes (comfy built-in and the Power Lora Loader)?

rgthree commented 1 week ago

I don't have all the nodes, so I can't run the workflow. But with the default Power Lora Loader, the numerical value is both model and clip strength. I see in the middle PLL it's set to "1.50" but on the one in the bottom, the model strength is "1.50" but the clip strength is "1.00" which would cause a difference, if those are the two portions you're comparing.

(You can also change the Power Lora Loader to have both model and clip strength separately from the node's Properties Panel)

gonzalu commented 1 week ago

Thanks... I found that after refreshing the session, it is much more consistent and results are 99% as expected. Much appreciated. It may just be a memory issue or something. I have noticed it a few times where the LoRAs don;t seem to trigger but after reloading ComfyUI or refreshing the browser session, it works as expected.

BTW, I wanted to say that the "hidden" features are FANTASTIC. The context menu options which I discovered accidentally are just fantastic, thank you.