Conditioning Adjust (Nukun)

There's a whole genre of ComfyUI nodes that promise to "enhance" your conditioning - and a good chunk of them are MLPs, attention layers, or calls into external packages you can't inspect. NukunConditioningAdjust is deliberately the opposite: a deterministic, preset-based conditioner with no hidden networks and no external code. Pick a preset, set a strength, get a report of what it did.

If that sounds boring, good. For Pony v7/T5 and SDXL/Pony v6 experiments, "boring and explainable" beats "mysterious and magical" every time - you can actually iterate when you know what the node did.

The presets

Five of them, in the info schema as preset:

• neutral_report_only - changes nothing, just reports. Use it as a baseline probe.
• literal_detail - pushes toward more literal, present-in-the-text detail. The README's go-to for Pony v7/T5 positives.
• soft_balance - a gentle balancing pass. The safest general starting point.
• contrast_pop - increases token contrast, more punch, riskier.
• negative_tamer - specifically for negative prompts, taming spikes. Use on Pony v7/T5 negatives.

model_profile can stay auto - it detects whether your conditioning is T5, SDXL CLIP, or SD1.5 CLIP and scales the adjustments conservatively (older CLIP conditionings get a milder detail push than T5). You can force t5, sdxl_clip, or sd15_clip if auto guesses wrong, but auto's the right default.

strength (0.0–2.0) scales the effect, and preserve_magnitude (default on) keeps the conditioning's length stable so your comparisons stay fair across the same seed. enabled is your bypass switch.

The recommended starting points

The README hands you a cheat sheet, and it's genuinely useful:

• Pony v7/T5 positive: literal_detail or soft_balance, auto profile, strength 0.5..1.0.
• Pony v7/T5 negative: negative_tamer, auto profile, strength 0.5..1.0.
• SDXL / Pony v6 positive: soft_balance, auto or sdxl_clip, strength 0.4..0.8.
• SD1.5 positive: soft_balance, auto or sd15_clip, strength 0.3..0.7.
• Diagnostics only: neutral_report_only, or use ConditioningAnalyzer instead.

How it works

Under the hood it does deterministic tensor ops - token contrast adjustment, norm smoothing, spike taming - chosen by preset and scaled by strength, with a magnitude-preservation pass. The report output tells you which profile, strength, and detected model family were applied, so the node doubles as its own documentation. That report is your debugging hook: if auto detected the wrong family, force the profile and re-run.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/OnekoSL/Nukun_ComfyUI_Nodes.git

Restart ComfyUI or install via ComfyUI Manager ("Nukun").

The honest verdict

This is an experimentation node, and it says so - deterministic, preset-based, built for "compare before/after on the same seed". It will not save a broken prompt or invent detail that isn't in your text; it gently reshapes how the conditioning pushes the model. For Pony v7/T5 users especially, where conditioning behavior is finicky and people get burned by opaque enhancers, having an explainable knob like this is quietly valuable. Start with soft_balance at a low strength, keep preserve_magnitude on, and compare against neutral_report_only before you dial anything up.