The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

@asomoza can you do a final review and test this out?

Nice, it works as intended and I really like the skip functionality, thank you for your work.

But I want to give my opinion here and open a discussion on about how is this implemented. Diffusers is used as a backend for UIs, for advanced users and also not for that advanced that knows all the diffusers naming for the blocks and layers, so I think we need to take this into consideration:

1. Doing this when loading the weights is restricting users and UIs that once the blocks are selected we can't change them, for that we have to reload the IP adapter. This is going to be specially noticeable when we use a lot of ip adapters, since even if we reload one, we need to reload all of them with the current implementation. I think we need to be able to change an IP adapter from a "full" configuration, a "style" configuration, a "composition" configuration or whatever we like on the fly without reloading.
2. Users will have to know this kind of names "up_blocks.0.attentions.1" to be able to use it and we already have the naming that @UmerHA implemented in Implements Blockwise lora #7352 which I like a lot more.

I really like to be able to use the same dict as the LoRAs for example.:

    {
        "down": {"block_1": [0.0, 0.0], "block_2": [0.0, 0.0]},
        "mid": 0.0,
        "up": {"block_0": 1.0, "block_1": [0.0, 0.0, 0.0]},
    },

on the other side, I really like the simplicity of this PR, if we go with the other implementation it will probably make the code a lot more complex.

I believe we can set different weights for each block just like LoRA dict, in this case, we don't need to reload modules. Will update soon.

Cool @haofanwang! What would be the mechanism to subtract the CLIP text embedding from the image embedding with diffusers? This is done here in the original implementation.

we will merge this one soon #7499
cool to rebase here? or is it easier the other way around?

Will rebase on #7499.

Is it possible to load multiple ip adapters at the same time?
ip-adapter + instantid
instantid + instantstyle

I would appreciate an example that shows how we could load/unload different adapters seperately or together.

Thanks!

@yiyixuxu @asomoza Our teammate has made a new PR. This PR will be closed soon.

Is it possible to load multiple ip adapters at the same time? ip-adapter + instantid instantid + instantstyle

I would appreciate an example that shows how we could load/unload different adapters seperately or together.

Thanks!

This is another problem. We can support it later.

Cool @haofanwang! What would be the mechanism to subtract the CLIP text embedding from the image embedding with diffusers? This is done here in the original implementation.

This can be natively achieved already. We will show how to do it once this PR merged.