Illustrious Noise Sampler (Nukun)

IllustriousXL and its finetunes are enormously popular for a reason - but they have a known quirk: fixed-seed, fixed-prompt runs can get stuck in a rut, producing variations that feel more like reshuffles than genuinely different images. The usual suspects (CFG, sampler choice) don't always fix it. One lever people don't reach for often enough is the initial noise itself. NukunIllustriousNoiseSampler is a compact sampler whose whole point is giving IllustriousXL more useful initial-noise variation than plain Gaussian.

It's the Illustrious-tuned member of the Nukun sampler family, sharing the advanced-sampler skeleton with Advanced Noise Sampler - same guider/sampler/sigmas/latent_image inputs, same zero-noise mode, same seed output, same preview_method control, same batch and noise-mask behavior.

What's different: composite noise

Instead of one noise_type, this node builds normalized composite noise - a weighted mix of Nukun noise components tuned for IllustriousXL. The controls:

• variation_mode - four recipes: balanced (default), texture, composition, wild. The names say what they favor: texture pushes toward detail/grain variation, composition toward larger structural forms, wild is everything turned up.
• variation_strength - final multiplier after normalization (1.0 default, up to 3).
• detail_bias - shifts the recipe's balance: lower values emphasize larger forms, higher values emphasize texture and micro-detail (default 0.35).
• noise_device - auto keeps CPU noise like ComfyUI core (reproducible); cuda when you want GPU noise and it's available.
• add_noise - off for zero-noise sampling, the rest of the standard set.

Outputs: output, denoised_output, and seed.

How to approach it

The README's framing is the right mental model: this is the first variation lever to try on IllustriousXL before the heavier one. The pack also ships a UNet Block Noise Patch that injects noise into the UNet itself - and the README explicitly says that's stronger and more model-sensitive, so the ordering matters. Try this sampler first, and only escalate to block noise if composite initial noise isn't enough.

A sensible tuning path:

1. Keep balanced at strength 1.0 and see if a fixed seed now produces more distinct outcomes.
2. If you want structure-level variety, nudge detail_bias down; if you want texture variation, nudge it up.
3. Save wild for when you're deliberately courting chaos and can review outputs.

Because it shares the family's preview_method - set latent2rgb for a lightweight per-step preview - you can bail early on a run that's clearly off-target instead of waiting for the full generation.

Installing it

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

Restart ComfyUI or install via ComfyUI Manager ("Nukun"). Runtime deps are the pack's standard set - numpy, Pillow, scipy, PyWavelets.

The honest verdict

This is a variation tool, not a quality tool - if your images are fine and you just want different content per run, this is a legitimate dial alongside seed and prompt edits. If your problem is quality (artifacts, bad anatomy), noise variation won't fix it; go back to the sampler, CFG, and model. And don't over-rotate on the settings - balanced at a modest strength captures most of the benefit, and the exotic modes exist for when you've exhausted the obvious ones.