Skip to content

Latest commit

 

History

History
143 lines (97 loc) · 4 KB

MAKING_A_MODEL_GUIDE.md

File metadata and controls

143 lines (97 loc) · 4 KB
Raw

Adapting this template for your model

This guide will help you adapt the cog-comfyui template repository for your own model.

If you haven’t used cog before or pushed a Replicate model, these are good starting guides:

Create a new repo from the template

Use https://github.com/fofr/cog-comfyui as a template to create a new repository

Prepare the template

After you have git cloned your new repository locally, including submodules, you should run the prepare_template.py script.

This will:

  • delete unnecessary files and directories.
  • overwrite the README.md with a blank file and header.
  • replace predict.py with example_predict.py

Run this script:

python scripts/prepare_template.py

Check what has been deleted and replaced before committing:

git status
git diff

Add your workflow

You should save the API version of your workflow as workflow_api.json.

It also helps to keep a copy of the UI version too, as workflow_ui.json.

Update the inputs to your model

predict.py is the entrypoint to your model. You can read about predict.py and the inputs you can use in the Cog documentation.

You'll end up with something like this:

def predict(
        self,
        prompt: str = Input(
            default="",
        ),
        negative_prompt: str = Input(
            description="Things you do not want to see in your image",
            default="",
        ),
        image: Path = Input(
            description="An input image",
            default=None,
        ),
        ...
    ) -> List[Path]:
        """Run a single prediction on the model"""
        ...

To make sure these inputs carry over to the workflow, you'll need to update the JSON object with new values. example_predict.py includes an example of this:

Within the predict method:

self.update_workflow(
    workflow,
    prompt=prompt,
    negative_prompt=negative_prompt,
    seed=seed,
)

And in the update_workflow method (in the Predictor class):

def update_workflow(self, workflow, **kwargs):
  # The node
  positive_prompt = workflow["6"]["inputs"]

  # Updating one of the nodes inputs
  positive_prompt["text"] = kwargs["prompt"]

  negative_prompt = workflow["7"]["inputs"]
  negative_prompt["text"] = f"nsfw, {kwargs['negative_prompt']}"

  sampler = workflow["3"]["inputs"]
  sampler["seed"] = kwargs["seed"]

Remove any custom nodes you do not need

To remove a custom node you should:

  • remove its corresponding entry from the custom_nodes.json file
  • optional: delete any corresponding helpers in custom_node_helpers
  • optional: delete any configs from custom_nodes_configs
  • optional: delete dependencies from cog.yaml

If you've already installed the nodes, make sure to also remove it from the ComfyUI/custom_nodes directory.

Add your own custom nodes

The simplest way to add new nodes is to:

  • add a new entry to the custom_nodes.json file, with the repo URL and the commit hash you want to use (usually the latest)
  • add any dependencies from the custom node’s requirements.txt to the cog.yaml file (if they are not already there)
  • rerun scripts/install_custom_nodes.py to install the new nodes

Some nodes will try to download weights on demand. You might want to avoid doing this from your Replicate model. The easiest fix is to make sure the downloaded weights are also pushed with your container. If you choose to do this, make sure to also update .dockerignore, which by default excludes weight from the container.

Running the model locally

You can run the model with defaults via cog:

cog predict

Or if you want to pass inputs:

cog predict -i prompt="something" -i image=@/path/to/image.jpg

Deploying your model to Replicate

Create a new model from https://replicate.com/create

Push your cog container to Replicate:

cog push r8.im/<username>/<model-name>