ComfyUI RALU Sampling Nodes

This is a standalone export of the ComfyUI custom nodes used for FLUX and SD3 RALU sampling experiments.

Quick Start / Installation

Clone the node package into an existing ComfyUI installation:

cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/Milkyawaway/comfyui-ralu-sampling.git ComfyUI-RALU-Sampling
cd /path/to/ComfyUI
python -m pip install opencv-python numpy typing-extensions
python main.py

Restart ComfyUI if it was already running. In the ComfyUI interface, load one of:

custom_nodes/ComfyUI-RALU-Sampling/examples/workflows/flux_ralu.json
custom_nodes/ComfyUI-RALU-Sampling/examples/workflows/sd3.5-ralu.json
custom_nodes/ComfyUI-RALU-Sampling/examples/workflows/ralu_flux_minimal_00001_.png

The PNG example contains embedded ComfyUI workflow metadata. Drag it directly onto the ComfyUI canvas to load the corresponding minimal FLUX RALU workflow.

The workflows expect model files in the standard ComfyUI model folders. The package can also be symlinked into custom_nodes/ when developing locally.

Main Files

• nodes_ralu_sampling.py: node definitions and ComfyUI mappings.
• ralu_engine.py: shared RALU helper logic.
• ralu_flux_runtime.py: FLUX RALU runtime.
• ralu_family_adapters.py: model-family adapters.
• sd3_two_stage_runtime.py: SD3 two-stage low/high-res runtime.
• ralu_helpers.py: utility functions and debug-info helpers.

Example Workflows

Example workflow files are in examples/workflows/:

• flux_ralu.json
• ralu_flux_minimal_00001_.png: generated minimal FLUX RALU output image with embedded drag-and-drop workflow metadata.
• flux_dev_t5fp16.json
• sd3.5-ralu.json
• sd3.5-lowres-test.json
• sd3.5-highres-debug.json

These workflows refer to common ComfyUI model filenames such as flux1-dev.safetensors, sd3.5_large.safetensors, clip_l.safetensors, clip_g.safetensors, t5xxl_fp16.safetensors, and ae.safetensors. Put the corresponding model files in the standard ComfyUI model folders before running.

Notes

The FLUX path is closer to the original mixed-token RALU structure. The SD3 path is a two-stage approximation and was mainly used to study grid artifacts and SD3's regular latent-grid limitations.