ComfyUI-FlowMatch-Advanced

Advanced Flow Matching and Noise Schedule scaling for DiT (Diffusion Transformer) models. This node replicates the precise inference logic from the lora-scripts ecosystem for Flux, Z-Image, and Qwen-Image.

Overview

Modern DiT models like Flux and Qwen use Rectified Flow or Discrete Flow Matching. A critical component of their success is the "Time Shift" logic, which adjusts the noise schedule based on the image's sequence length (resolution).

Without this shift, images at higher resolutions (e.g., 2048px) often suffer from structural collapse or lack of detail because the noise schedule is too "steep." This node automates that math for you.

Features

• Resolution-Aware Shifting: Automatically calculates the mu value based on your target resolution, replicating the $0.5 ightarrow 1.15$ shift range used in Black Forest Labs' Flux and Lumina-Next.
• DiT Optimized Presets: Specialized math for:
  • Flux: MM-DiT sequence scaling.
  • Z-Image: (Lumina-Next/Flag-DiT) Discrete flow scaling.
  • Qwen-Image: (Lumina-Qwen) LLM-backbone specific shifting.
• V-Prediction Enforcement: Automatically patches the model to v_prediction mode to prevent washed-out/grey images.
• Universal Compatibility: Works with both Load Checkpoint and Load Diffusion Model workflows.

Installation

1. Navigate to your ComfyUI/custom_nodes/ directory.
2. Clone this repository:

   git clone https://github.com/supSugam/ComfyUI-FlowMatch-Advanced
   

3. Restart ComfyUI.

Usage

Node Parameters

• model: The DiT model you want to patch (from Load Checkpoint or Load Diffusion Model).
• model_type: Choose the architecture (flux, z-image, or qwen).
• shift: The manual shift value (used if resolution_aware is off, or as the max boundary for Qwen).
• resolution_aware: (Recommended) When enabled, it ignores the manual shift and calculates the optimal trajectory based on width/height.
• width/height: Your target generation dimensions (used for the mu calculation).

Recommended Workflow

1. Load Model: Load your Flux/Z-Image/Qwen model.
2. Apply FlowMatch: Connect the MODEL output to the DiT FlowMatch node.
3. Connect Sampler: Connect the patched MODEL to your KSampler or SamplerCustom.
4. Sampler Settings:
   • Sampler: euler
   • Scheduler: normal or sgm_uniform
   • Denoise: 1.0 (for full generation)

Why this node?

In the original lora-scripts and official inference implementations, the noise schedule isn't static. It follows this formula:

\sigma = \frac{shift \cdot t}{1 + (shift - 1) \cdot t}

And for Flux, the shift (mu) is determined by:

mu = lerp(0.5, 1.15, \text{sequence\_length})

This node brings that exact mathematical precision to ComfyUI, ensuring your images look exactly as they would when generated with the specialized training scripts.