well, it's not on lllasvials list, so might be just not supported atm on auto1111. you can ask the devs on github for a clearer answer about it, perhaps they are already working on it.

you change the preprocessor to "none" for it to work, though it will have close to no effect

it is rather weak and has to be used differently than the 1.5 inpainting cn. No preprocessor and the inputimage has to be completely white where changes should happen. I'd recommend to use it with regular inpainting and use it twice (or even more often, depending on how much control you need) for one inpaint with different areas marked for inpainting.

Start with the smallest area you want to change and with a weaker cn strength (like 0.3). The controlnet will try to keep everything as it that is not marked as white. With every further controlnet you can increase the size of the marked area and play with the cn strength. The biggest/last area should be as large as the inapint area itself. Play with different cn strengths for the different areas and see for yourself. Compared to sd1.5 inpainting models it's a lot more work (easier with the photoshop plugin and reusable layers), but you have a lot more control and don't loose colors/sharpness when inpainting multiple times in a row. I hope someone will release a stronger version of this, but even this one is quite usable (with denoising strength of 1.0, even with latent noise)

no, they should work with the names they have. where do you get that from?

@eurotaku  in your see more section you wrote

As files with the extension .pth can't be uploaded the ip-adapter.pt from h94 has to be renamed manually after downloading.

@Malessar ah, yes, you're right. i reuploaded it as .bin and .safetensors when i added the other ip-adapters, as the .pth extension is only necessary for sd1.5 afaik. i changed the description, thx for mentioning it :)

er, that depends on what exactly you want to do with it. probably one of the most fun controlnets to play around with would be the ip-adapter, being able to transfer styles, compositions and poses from one image to another. just experiment with its weight and your prompt to create interesting results.

@eurotaku thank you for the reply, are there any examples to see what are the differences between the models? I understand every model has its on purpose and the authors probably trained them all in a different way, just wanted to see what's the difference e.g. between kohya's candy and bdsqlz candy

btw ...

where ic can get an idea how to handle the models (oc controllnet i know) and which is good for what ?

Have a look at the galleries displaying the different input images they take/the preprocessors creates from the original input e.g. https://github.com/Mikubill/sd-webui-controlnet?tab=readme-ov-file#gallery

The early/basic controlnets were trained on classical computer vision tasks, like edge detection images (canny), depth estimation as grayscale (depth, normal, ...), or pose detection images (openpose). They are good if you want to have a finegrained strict control of how your image should be composed (e.g. where edges should be, where people should be placed)

The more advanced ones like the ipadapters (https://github.com/tencent-ailab/IP-Adapter?tab=readme-ov-file#sdxl_10) are more abstract and try to transfer more complex concepts. E.g. the ip-adapter-plus-face will try to replicate a given face (or rather its features) within the capabilities of the used SD model. Playing arround with the controlnet strength and the (de-)activation steps can be helpful to finetune your output and balance between the controlnets and the prompt.