Conditioning Parameters v2

This is the box in the Searge SDXL workflow that controls how your prompts get turned into conditioning - the strength scales, the aesthetic scores, and an optional noise pre-conditioning step. It's the "Conditioning Parameters v2" node. Same pattern as the rest of the Searge inputs: it doesn't encode anything itself, it just records numbers into the data-stream bus that the Magic Box's conditioning stage reads back out.

If you've only ever used a plain CLIP Text Encode, most of these controls will be unfamiliar, and that's fine - the defaults are tuned and you can run the whole workflow without touching them. This box is for when you want finer control over how hard the model listens to your positive versus negative prompt, and how SDXL's aesthetic-score mechanism is set.

How it works

SDXL's conditioning carries more than just your prompt text. It also encodes a target resolution and an "aesthetic score" - a number the base and refiner were trained to associate with image quality. Searge exposes those as knobs. On top of that it lets you scale positive and negative conditioning independently, and optionally mix a little Gaussian noise into the conditioning as a "pre-condition" for variety.

The inputs and outputs that matter

There's a lot here, but only a few you'd realistically adjust:

• positive_conditioning_scale (default 1.5) and negative_conditioning_scale (0.75) - how strongly each prompt pushes. The most useful pair. Nudge the negative up if unwanted stuff keeps appearing.
• positive_aesthetic_score (6) and negative_aesthetic_score (2.5) - SDXL's aesthetic-score conditioning. High for what you want, low for what you don't. The defaults are the standard values and rarely need changing.
• base_/refiner_/target_conditioning_scale (all default 2) - coarser scales feeding the base pass, refiner pass, and target. Leave at defaults unless you're deliberately experimenting.
• precondition_mode (none / gaussian) and precondition_strength (0.1) - optional noise injected into the conditioning. none keeps things clean and deterministic.

Output is the single data (SRG_DATA_STREAM) stream.

How to install it

Manager: search SeargeSDXL, install, restart. Manual install (opencv first, it's a hard dependency):

python -m pip install opencv-python
cd ComfyUI/custom_nodes
git clone https://github.com/SeargeDP/SeargeSDXL.git

Restart ComfyUI. No model files needed for this node.

Common issues & troubleshooting

Aesthetic score isn't a magic quality dial. People see "aesthetic score 6" and crank it to 10 expecting prettier images. It doesn't work like that - it's a conditioning signal the model was trained against, and the 6 / 2.5 split is the standard, well-behaved setting. Extreme values tend to distort more than they improve.

Negative prompt feels weak. SDXL doesn't respond to the wall-of-negatives habit that SD 1.5 users bring over; shorter negatives generally work better. If yours genuinely isn't biting, raising negative_conditioning_scale is the lever, but trim the negative prompt itself first.

Only touch this after the basics. The defaults in this box are considered good. If you're new to the Searge workflow, get the model selector and prompts working and leave conditioning parameters alone - it's a refinement tool, not a first stop.