ComfyUI Extension: 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.
Custom Nodes (0)
README
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
-
Clone this repository to your ComfyUI custom nodes directory:
cd ComfyUI/custom_nodes/ git clone https://github.com/opparco/ComfyUI-WanLightx2vScheduler
-
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 fromsteps
: 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 samplingpositive
: Positive conditioningnegative
: Negative conditioninglatent_image
: Input latentsampler_name
: Sampler algorithmsigmas
: Custom sigma schedulecfg
: CFG scale (0.0-100.0, default: 1.0)start_at_step
: Starting step (default: 0)end_at_step
: Ending step (default: 4)add_noise
: Whether to add noise (default: True)noise_seed
: Random seed for noise generation
- Outputs:
output
: Final sampled latentdenoised_output
: Denoised output (when available)
Usage Example
Basic Workflow:
- Load your Wan2.2 Lightx2v model
- Add
WanLightx2vSchedulerBasic
node - 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
- Connect sigmas output to
KSamplerAdvancedPartialSigmas
- Configure sampler parameters as needed
Contributing
Contributions are welcome! Please feel free to submit issues and enhancement requests.