ComfyUI Inference Sampler

A ComfyUI custom node implementing Pareto-weighted adaptive sampling that blends historical gradient context with the current diffusion step.

Installation

1. Clone or copy this folder into ComfyUI/custom_nodes/
2. Restart ComfyUI
3. The Inference Sampler node will appear under the sampling category

Parameters

Parameter	Default	Range	Description
heterodyne_strength	0.20	0.0–1.0	Mix of current gradient vs. historical context. Higher = more past influence.
gradient_strength	0.40	0.0–1.0	Momentum from the previous step. Higher = smoother path.
context_weight	0.80	0.0–1.0	Final blend back toward the raw current gradient. Higher = more current.
alpha	0.71	0.1–10.0	Pareto shape. Lower = diffuse (many steps matter). Higher = concentrated (only aligned steps matter).
gradient_history_length	8	2–20	How many past gradients to retain for context.
min_alignment_threshold	0.80	0.01–1.0	Minimum alignment value used as the Pareto distribution offset.

How It Works

Each step:

1. Computes cosine similarity between the current gradient and each stored historical gradient
2. Applies a Pareto PDF to weight historical gradients by their alignment score
3. Blends the Pareto-weighted context with momentum from the previous step (heterodyne_strength, gradient_strength)
4. Blends that combined direction back with the raw current gradient (context_weight)
5. Steps forward and stores the gradient for the next iteration

Requirements

• ComfyUI 0.17.0+
• Python 3.10+
• PyTorch

Disclaimer

Obfuscated because I felt nostalgic

License

Arc-9 Open Royalty Agreement 1.4