Wan2.2 Lightx2v Scheduler for ComfyUI

A custom ComfyUI node package designed specifically for Wan2.2 Lightx2v models to fix the "burnt-out" look, over-sharpening, and abrupt lighting shifts through proper denoising trajectory alignment.

Problem & Solution

The Issue

When using Wan2.2 with the lightx2v LoRA, users commonly experience:

• "Burnt-out" appearance with excessive contrast
• Over-sharpening artifacts
• Abrupt lighting shifts between frames

The Solution

This package generates custom sigmas that recreate the exact denoising trajectory the LoRA was trained on, ensuring consistent results across different step counts.

Installation

1. Clone this repository to your ComfyUI custom nodes directory:

   cd ComfyUI/custom_nodes/
   git clone https://github.com/opparco/ComfyUI-WanLightx2vScheduler
   

2. Restart ComfyUI

Nodes Included

WanLightx2vSchedulerBasic Recommended

• Purpose: Precise sigma scheduling with theoretical accuracy
• Inputs:
  • steps: Number of sampling steps (1-10000, default: 4)
  • sigma_max: Maximum sigma value (Use 1.0 for theoretical accuracy)
  • sigma_min: Minimum sigma value (Use 0.0 for theoretical accuracy)
  • shift: Time shift parameter (0.1-100.0, use 5.0 for lightx2v)
• Output: SIGMAS tensor for custom sampling

WanLightx2vSchedulerBasicFromModel

• Purpose: Automatic sigma scheduling using model parameters (may not match theoretical values)
• Inputs:
  • model: The model to extract sigma parameters from
  • steps: Number of sampling steps (1-10000, default: 4)
  • shift: Time shift parameter (0.1-100.0, default: 5.0)
• Output: SIGMAS tensor for custom sampling
• Note: Use Lightx2vSchedulerBasic with sigma_min=0.0, sigma_max=1.0 for best results

KSamplerAdvancedPartialSigmas

• Purpose: Advanced sampler supporting custom sigma schedules and partial step execution
• Inputs:
  • model: Model for sampling
  • positive: Positive conditioning
  • negative: Negative conditioning
  • latent_image: Input latent
  • sampler_name: Sampler algorithm
  • sigmas: Custom sigma schedule
  • cfg: CFG scale (0.0-100.0, default: 1.0)
  • steps: Number of steps to execute (1-10000, default: 4)
  • add_noise: Whether to add noise (default: True)
  • noise_seed: Random seed for noise generation
• Outputs:
  • output: Final sampled latent
  • denoised_output: Denoised output (when available)

Usage Example

Basic Workflow:

1. Load your Wan2.2 Lightx2v model
2. Add WanLightx2vSchedulerBasic node
3. Set parameters:
   • sigma_min: 0.0 (for theoretical accuracy)
   • sigma_max: 1.0 (for theoretical accuracy)
   • shift: 5.0 (matches LoRA training trajectory)
   • steps: 4, 8, 16, or 20
4. Connect sigmas output to KSamplerAdvancedPartialSigmas
5. Configure sampler parameters as needed

Contributing

Contributions are welcome! Please feel free to submit issues and enhancement requests.