[Feature Request]: Replicate support

[platform-kit]

Is there an existing issue for this?

• I have searched the existing issues and checked the recent builds/commits

What would your feature do?

Hi! Great work with this fork.

Putting in a request for adding replicate.com support.

Proposed workflow

It's as simple as adding two files - a cog.yaml that specifies underlying pip packages and other environment details (CUDA version etc)

https://github.com/mrhan1993/Fooocus-API/blob/main/cog.yaml

and a predict.py where the model is loaded and the input types are specified

https://github.com/mrhan1993/Fooocus-API/blob/main/predict.py

You would just have to copy these two files to the repo and then modify to add support for whatever changes you've made to enable InstantID, etc.

Additional information

Further info on the Cog package, which is a wrapper for docker, at https://cog.run

[platform-kit]

@machineminded any chance of getting this on the roadmap? I am happy to implement it myself if you can point out what inputs/variables need to be passed.

[machineminded]

I haven't had a chance to look at this yet. But please, if you want to integrate this feel free. What inputs are you looking for? The inswapper stuff?

[platform-kit]

@machineminded

I'll need to know what packages you added to the branch so I can list them in the requirements in
cog.yaml as well as what new input fields you're now passing to the API to enable/disable new features.

In a cog model (Replicate's format) you do all the prediction in a single file https://github.com/mrhan1993/Fooocus-API/blob/main/predict.py

You specify the possible input fields in def predict (line 50 in the above) which can then you can access when calling the underlying model code (fooocus, in this case) in the subsequent code.

For example in the above code they declare an input variable inpaint_input_image then use it on line 259, to pass to the Fooocus API like so:

     params = ImageGenerationParams(
            prompt=prompt,
            negative_prompt=negative_prompt,
            style_selections=style_selections_arr,
            performance_selection=performance_selection,
            aspect_ratios_selection=aspect_ratios_selection,
            image_number=image_number,
            image_seed=image_seed,
            sharpness=sharpness,
            guidance_scale=guidance_scale,
            base_model_name=base_model_name,
            refiner_model_name=refiner_model_name,
            refiner_switch=refiner_switch,
            loras=loras,
            uov_input_image=uov_input_image,
            uov_method=uov_method,
            upscale_value=uov_upscale_value,
            outpaint_selections=outpaint_selections_arr,
            inpaint_input_image=inpaint_input_image_dict,
            image_prompts=image_prompts,
            advanced_params=advanced_params,
            inpaint_additional_prompt=inpaint_additional_prompt,
            outpaint_distance_left=outpaint_distance_left,
            outpaint_distance_top=outpaint_distance_top,
            outpaint_distance_right=outpaint_distance_right,
            outpaint_distance_bottom=outpaint_distance_bottom,
            save_meta=True,
            meta_scheme='fooocus',
            save_extension='png',
            save_name='',
            require_base64=False,
        )

IMO the best course of action would be to do a minimal implementation of InstantID.

What would be the minimal input needed to invoke a generation with InstantID? I'd imagine you're using some kind of boolean like enable_instantid and an image field like instantid_input_image?

What are those fields and how do you pass them to the underlying Fooocus model?