Files

49 lines
3.1 KiB
Markdown

## TT SDXL Workflow Settings *TT_SdxlWorkflowSettingsNode*
*This description was generated with AI assistance. If you spot any errors, please create an issue on GitHub.*
Configures global workflow parameters for SDXL pipelines. Outputs a workflow config object alongside individual parameters for direct node
connections.
### Inputs
| Parameter | Type | Required | Description |
|-------------------|-------|----------|--------------------------------------------------------|
| `seed` | INT | Yes | Generation seed |
| `steps` | INT | Yes | Number of sampling steps, default 30 |
| `cfg` | FLOAT | Yes | Guidance scale, default 5.0 |
| `sampler_name` | COMBO | Yes | Sampler algorithm |
| `scheduler` | COMBO | Yes | Noise schedule |
| `ascore_positive` | FLOAT | Yes | Aesthetic score for positive conditioning, default 9.0 |
| `ascore_negative` | FLOAT | Yes | Aesthetic score for negative conditioning, default 6.0 |
| `width` | INT | Yes | Image width in pixels |
| `height` | INT | Yes | Image height in pixels |
| `target_width` | INT | No | CLIP target width (advanced) |
| `target_height` | INT | No | CLIP target height (advanced) |
### Outputs
| Parameter | Type | Description |
|-------------------|-----------------|-------------------------------------------|
| `WORKFLOW_CONFIG` | WORKFLOW_CONFIG | Workflow config object for context nodes |
| `SEED` | INT | Generation seed |
| `STEPS` | INT | Number of sampling steps |
| `CFG` | FLOAT | Guidance scale |
| `SAMPLER_NAME` | STRING | Sampler algorithm name |
| `SCHEDULER` | STRING | Noise schedule name |
| `ASCORE_POSITIVE` | FLOAT | Aesthetic score for positive conditioning |
| `ASCORE_NEGATIVE` | FLOAT | Aesthetic score for negative conditioning |
| `WIDTH` | INT | Image width |
| `HEIGHT` | INT | Image height |
| `TARGET_WIDTH` | INT | CLIP target width |
| `TARGET_HEIGHT` | INT | CLIP target height |
### Usage
Connect `WORKFLOW_CONFIG` to context-based nodes. Use individual outputs for direct connections to sampler and encoder nodes. Set `target_width` and
`target_height` to match the original training resolution for best results — defaults to image dimensions if not set.
---
[View on GitHub](https://github.com/tenser-tensor/ComfyUI-TenserTensor)