|
|
--- |
|
|
license: apache-2.0 |
|
|
language: |
|
|
- en |
|
|
tags: |
|
|
- LoRA |
|
|
- Negative |
|
|
- SDXL |
|
|
- illustrious |
|
|
- Pony |
|
|
--- |
|
|
|
|
|
mg_7lambda_negative |
|
|
Short description: a LoRA for suppressing unwanted visual traits (negative bias). Tuned and tested with SDXL (ComfyUI). |
|
|
|
|
|
Note: perceived strength depends on the base model and prompt. Start low and increase gradually. |
|
|
|
|
|
Files |
|
|
|
|
|
How to Use (ComfyUI) |
|
|
|
|
|
Download the file and place it in ComfyUI/models/loras. |
|
|
Classic graph: add a Load LoRA node, select the file, set strength_model and strength_clip. |
|
|
MagicNodes CombiNode: enable a slot (use_lora_X = True) and pick this LoRA in the dropdown. |
|
|
Recommended Settings (starting points) |
|
|
|
|
|
strength_model: -0.40–1.00 |
|
|
|
|
|
strength_clip: 0.20–0.40 |
|
|
|
|
|
If the effect is too strong, lower both by 0.05–0.10 per step. |
|
|
|
|
|
For more controlled results: try slightly lower CLIP (e.g., 0.35) and slightly higher model (e.g., 0.55). |
|
|
|
|
|
Prompt Tips |
|
|
|
|
|
Put the negative prompt: (super-wrong:1) |
|
|
|
|
|
|
|
|
Known Limitations |
|
|
|
|
|
Primarily tuned for SDXL; other bases may need different strengths. |
|
|
|
|
|
Very high values can “wash out” details. |
|
|
|
