WEEK 8 · LESSON 10 · TOPIC 2 OF 6 · How It Works

How It Works

Lesson progress

0% Complete

SECTION 1 OF 1

In this topic · 1 sections
  1. Overview

Overview

Examples of external tools with ComfyUI nodes include:

Mimic Motion: Creates a dancing or motion-based sequence from a single character image.

ToonCrafter: Accepts a starting frame and ending frame and generates an anime sequence between them.

LivePortrait: Accepts a face image and a driving motion clip, then maps the clip's expressions and head movements onto the input face.

AnimateDiff: Supports animation generation from images.

Each integration has its own graph, but the operational pattern is consistent: install the custom node, satisfy its dependencies and model requirements, then connect its inputs and outputs into the ComfyUI pipeline.

For InstantID, the setup proceeds in several layers:

  1. Install the native InstantID support package through Custom Nodes Manager.
  2. Restart ComfyUI.
  3. Allow startup to install dependencies, including InsightFace and ONNX Runtime GPU for the CUDA setup.
  4. Download the AntelopeV2 InsightFace model archive.
  5. Create ComfyUI/models/insightface/models if it does not exist.
  6. Extract the AntelopeV2 folder inside that directory.
  7. Use Model Manager to install the InstantID IP-Adapter model and the InstantID ControlNet model.
  8. Refresh ComfyUI so the new assets become selectable.

The Apply InstantID node is inserted before the KSampler. It receives and modifies the model and conditioning used during sampling. The graph connects:

The checkpoint model to Apply InstantID, then the modified model to the KSampler.

The positive and negative prompt conditioning through the InstantID stage and into the KSampler.

A loaded reference-face image.

An InstantID ControlNet selected with ControlNet Loader.

InstantID Face Analysis, with CUDA selected as the provider in the GPU example.

The InstantID model loaded from the installed IP-Adapter file.

The remaining text-to-image graph stays familiar: an Empty Latent Image enters the KSampler, the result is VAE-decoded, and the image is previewed.

InstantID exposes three important controls:

Weight: How strongly the reference face influences the final image.

Start: The fraction of the denoising process at which the identity influence begins.

End: The fraction at which it stops.

If the KSampler uses 20 steps and End is set to 0.5, the identity influence stops around step 10. If End is 1.0, it remains active through the full denoising process.

A realistic police-officer prompt can use a reference face, then shift to a watercolor-painting prompt after removing painting from the negative prompt. The face remains recognizable across the style change. InstantID can also maintain face fidelity across different viewing angles from a single reference image.