how to import a custom pipline?

[micklexqg]

As some piplines, such as DiffusionPipeline and StableDiffusionPipeline, are merged into diffusers after installation, so how to import a new custom pipline, such as StableDiffusionXXXPipeline which is newly coded.

[asomoza]

Hi, when you are referring to custom pipeline which is newly coded, you mean one in the examples/community directory right? Also you mean when the pipeline is merged into main but we still haven't done a release?

The answer is that you will need to install diffusers from source and then manually import the pipeline, alternatively you can just download the pipeline and also manually import it. Once we do a release, you can then use the custom_pipeline argument.

[micklexqg]

Hi, when you are referring to custom pipeline which is newly coded, you mean one in the examples/community directory right? Also you mean when the pipeline is merged into main but we still haven't done a release?

The answer is that you will need to install diffusers from source and then manually import the pipeline, alternatively you can just download the pipeline and also manually import it. Once we do a release, you can then use the custom_pipeline argument.

I have installed diffusers and try to import a newly coded pipline which is "StableDiffusion3ControlPipeline" here. I tried with two ways but failed.
as follows:

the import line code is gray.
I also tried manually but still failed, as follows:

the import line code is white.

[asomoza]

they key part here is I have installed diffusers, you don't specify how you installed diffusers. Also that pipeline doesn't exist in our repository so I'm guessing that with newly coded you mean something that you coded?

If that's not the case, and you want to import the diffusers SD3 controlnet pipeline, you have to do it like this:

from diffusers import StableDiffusion3ControlNetPipeline

[micklexqg]

Thanks, the given demo such as train_dreambooth_sd3 runs ok, and i make some change in the demo, at present the training runs ok, i need add a new pipline for validation, as shown in the project structure, i add a StableDiffusion3ControlNetPipeline, but failed to import it using "from diffusers import StableDiffusion3ControlNetPipeline" , need I import the new custom pipeline using relative or absolute path?

[asomoza]

you have several options, then one you're trying to do is the more complex one because you need to change the diffusers source code. The only way to do from diffusers import CustomPipeline is if you insert the appropiate imports in all the __init__.py files from the pipeline directory all the way to the base src directory, this will tie your pipeline to your specific custom version of diffusers.

What I recommend you do so you can use it everywhere, is to upload it to the hub, like I did here for this custom SDXL pipeline and then just load it like:

pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
    "SG161222/RealVisXL_V5.0_Lightning",
    torch_dtype=torch.float16,
    vae=vae,
    custom_pipeline="OzzyGT/pipeline_sdxl_fill",
    controlnet=controlnet_model,
    variant="fp16",
).to("cuda")

another option is to copy your pipeline to the root of the project and import it directly like:

from .StableDiffusion3ControlPipeline import StableDiffusion3ControlPipeline

but for both cases you will need to change the imports inside your pipeline to match where they are since you moved the file.

The final option you have is to import from the file instead of the directory like you're doing, like for example:

from diffusers.pipelines.stable_diffusion_3.StableDiffusion3ControlPipeline import StableDiffusion3ControlPipeline 

All of this makes me think that you are having trouble not with diffusers but with the python import system, I suggest you start with some basic project and learn how they work so you don't have issues in the future and in the case you don't want to, I recommend you use the first option.

[micklexqg]

Thanks a lot