FlowMatch Euler Discrete Scheduler for ComfyUI

9 steps, big res, zero noise.

FlowMatchEulerDiscrete seems not exposed in ComfyUI, but it is what the official Z-Image demo in diffusers use.

So:

• I am exposing it in the scheduler section for you to use within KSampler.
• On top I provide a node, experimental, to configure the scheduler for use with CustomSampler and play with.

In short...if you want sharper and noise free images, use this!

Installation

• use comfy ui manager (search erosDiffusion or ComfyUI-EulerFlowMatchingDiscreteScheduler)

or

• git clone https://github.com/erosDiffusion/ComfyUI-EulerDiscreteScheduler.git in your custom nodes folder.

Example output (more below)

What you get

• one new scheduler FlowMatchEulerDiscreteScheduler registered in the KSampler
• a custom node that exposes all parameters of the FlowMatchEulerDiscreteScheduler which Outputs SIGMAS for use with SamplerCustom node.

Usage

• Simple: select the FlowMatchEulerDiscreteScheduler in the default workflow from ComfyUI and run.
• Advanced/experimental:
  1. Add FlowMatch Euler Discrete Scheduler (Custom) node to your workflow
  2. Connect its SIGMAS output to SamplerCustom node's sigmas input
  3. Adjust parameters to control the sampling behavior, you have ALL the parameters to play with.

Tech bits:

• https://huggingface.co/docs/diffusers/api/schedulers/flow_match_euler_discrete
• https://huggingface.co/Tongyi-MAI/Z-Image-Turbo/blob/main/scheduler/scheduler_config.json

Find this useful and want to support ?

Buy me a beer!

More examples:

Changelog

1.0.7

• nunchaku qwen patch fix, tiled diffusion patch fix users reported issues with dimensions not being handled correctly, this should fix it.

1.0.6

• updated example
• updated pyproject deps (diffusers)

1.0.5

• remove bad practice of forking diffusers install on error (requirements.txt and does not rollback your diffusers if available)

1.0.4

• add start and end step by Etupa, with some fixes (can be used for image to image or restart sampling)

1.0.3

• node publish action

1.0.2

• changed the device management in the custom scheduler node to be on gpu (cuda)
• removed flash attention node dependency from the custom scheduler node
• removed flash attention node from init
• added mit licensing