[Feature request] Capability to load models from other nodes, not only from disk

[jovan2009]

First of all, thank you for these awesome nodes!

Right now I'm playing with the LoRA extract nodes, I like very much the "skip" mismatch mode. When I extract LoRA's from Wan 2.2 models, using "skip" avoids getting console spam about missing weights when I use the resulted LoRA's.

About the feature request:
What I would like to be able to do is to to extract not only from [Model A - Model B], but also from [(Model A + Lora X) - (Model B + Lora Y)]. That would mean to be able to do what I can currently do with LoraExtractKJ:

Something like I simulated here, if that would be possible:

I know that technically I can do what I ask right now but would mean to merge models with LoRA's first and save them to disk, not on the fly.

Thanks in advance!

[silveroxides]

This would break the internal principle of these custom nodes as they are meant to have low impact on memory. Adding arbitrary model object input will break this.

[jovan2009]

Thanks for your response. I wasn't aware of the implications.

I was ready to close this issue but I got the idea that what I want doesn't necessarily need what I proposed in my opening post. Seeing the functionality of those two nodes I wonder if they can combined, so the Lora Extract Node can have, as inputs, instead of Model A and Model B, Model A + Lora A and Model B + Lora B?

I apologize for my ignorance if my new idea has again the same downsides, I'm obviously not a developer. I just discovered that your custom node make possible to experiment with various variants of Wan 2.2 models that are not 1:1 compatible with each other (like Wan 2.2 Camera Fun and Wan 2.2 I2V distilled). But you can get interesting effects by adding and subtracting intermediate Lora's and what stops me to experiment further is the current need to save Model A + Lora A and Model B + Lora B to disk. And I'm very low on disk space (and I think that my SSD hates me already from the abuse I already did to it).

Thanks in advance!

[silveroxides]

@jovan2009 This adds too much complexity for little gain. Handling LoRa layer naming, layer matching, shapes and all this is hell enough as is. Consider waiting with doing such experimental merges until you have expanded your storage.

[jovan2009]

Thanks, understood.
Closing now.