convert_ic-light.py will take a regular 1.5 ckpt and the ic light model and combine them into a single diffusers format ip2p model. Then use the normal convert_diffuser_to_original_stable_diffusion.py script from diffusers to convert it back to a ckpt for comfy.
The results are slightly different but I think that may be due to ELLA vs normal CLIP. The lighting latent might also need to be scaled depending on the light pattern/colors, not sure yet why there's a difference there.