Pony V7 Noise Sampler (Nukun)

The boring truth about initial noise is that most of the time gaussian at strength 1.0 is exactly right, and anyone selling you "special noise for your model" is usually selling nothing. This node is the exception that earns a look, because Pony v7 is not like other models. It's an AuraFlow-based T5 model that shipped to mixed reception and, per the KB, is "very conservative in terms of VRAM usage (GGUF)" but notoriously twitchy about quality and structure. Its community found that tuned, lower-energy composite noise - not more noise, less - stabilizes the first pass and the refine pass in ways plain Gaussian doesn't.

The mechanism: instead of one noise type, the node builds normalized composite noise from multiple components - a Gaussian-dominated first stage and a violet-leaning refine stage - and scales the result with noise_strength. It's a drop-in SamplerCustomAdvanced-style sampler: you feed it guider, sampler, sigmas, latent_image, and it handles the initial noise generation internally, returning the sampled output latent, a denoised_output, and the seed.

The settings that matter

• v7_profile - the recipe. stage1_gaussian (default, ~0.55 strength) for the first full pass, stage2_violet (~0.50) for the refine/HiResFix pass. balanced, soft, and graphic are alternatives; detail_bias shifts the recipe toward larger forms (low) or micro-detail (high).
• add_noise - disable to sample with zero noise, which is how you chain a two-stage pass: first stage adds noise, second stage runs with add_noise off.
• noise_device - auto keeps ComfyUI-core-like CPU noise (reproducible); cuda uses CUDA when available, falling back to CPU. Leave it on auto unless you're chasing a specific reproducibility change.
• preview_method - default follows the global queue setting, latent2rgb gives a lightweight per-step preview, taesd uses TAESD assets, none disables. On low VRAM, latent2rgb lets you cancel early when a run is obviously off-target.

The node keeps the same noise_mask, batch_index, and nested-latent behavior as its siblings in the pack, so it slots into existing advanced workflows without surprises.

How to use it

This is an advanced-sampler-format node, so you need the explicit graph: CFGGuider → KSamplerSelect → BasicScheduler → this node → VAE decode. The typical Pony v7 two-pass pattern: first pass with stage1_gaussian at strength ~0.55, then a refine pass (usually a HiResFix-style second sampler) with stage2_violet around 0.50 and add_noise off. The defaults are tuned for exactly that flow, which is refreshing - most sampler nodes ship neutral defaults and leave you to guess.

Installing it

Same pack install as everything else:

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

Restart ComfyUI, or install via Manager (search "Nukun"). The composite-noise machinery is self-contained in the pack (numpy/scipy/PyWavelets from requirements.txt); it doesn't need ComfyUI-sonar even though the expanded profiles are Sonar-inspired.

Common issues

The trap is over-strengthening. These profiles are normalized composites, so noise_strength above ~1.0 produces genuinely aggressive initial noise that Pony v7 punishes with artifacts - the node's own guidance (0.55 / 0.50) is lower than the classic 1.0 Gaussian, and that's intentional, not a bug. If you're coming from Illustrious Noise Sampler or Universal Noise Sampler, remember the Pony recipes are tuned for lower energy; porting your old strength settings over usually looks wrong. And if you see zero-noise behavior you didn't ask for, check add_noise - it defaults to true, but the toggle is right next to everything else.