Glitch Sigmas

Every sampler in ComfyUI runs down a noise schedule - the sequence of sigma values that decides how much noise gets removed at each step. A Karras schedule descends smoothly, which is why it's the default for so many models. GlitchSigmas takes that smooth curve and adds random multiplicative wobble to every step. The schedule no longer descends evenly, the denoiser gets confused about how noisy the image is supposed to be, and the tonal range warps into that granular, burned-print look. Same prompt, same composition - wrong physics. That's the whole trick.

What it does, precisely

The node is MODEL → SIGMAS. It builds a base scheduler curve through ComfyUI's own calculate_sigmas (using your model's model_sampling, so it's model-aware rather than a hardcoded table), then applies jitter: each sigma gets multiplied by (1 + jitter × noise) where the noise is seeded and per-step.

Two details keep it from being a crash generator instead of a glitch tool:

• The terminal sigma is re-pinned to the base value. If the last step doesn't reach the end of the schedule, denoising never completes and you get a half-cleaned image. Re-pinning means the loop still finishes.
• Intermediate sigmas are clamped positive. A negative sigma in the middle of a run is how you get NaNs and black squares.

The jitter slider runs 0–1. At 0 the node is just a plain, unperturbed scheduler builder - a useful reference point and a harmless no-op. At 0.50 you're at the README's "schedule-burn" preset: composition stays coherent (because prompt influence isn't touched) but the tonal range is visibly warped. The effect reads like a print that got overdeveloped, not like noise added on top.

The wiring catch

This is the one thing beginners trip on. GlitchSigmas outputs a SIGMAS tensor, and the only stock sampler that accepts an explicit SIGMAS input is SamplerCustomAdvanced. Plain KSampler computes its own schedule internally - there's no sigmas socket on it.

So the right graph is the full advanced pathway:

CheckpointLoaderSimple → AttentionDropoutPatcher (optional) → CFGGuider
EmptyLatentImage → SamplerCustomAdvanced → VAEDecode
                    (noise = RandomNoise, sampler = SamplerEulerAncestral,
                     sigmas = GlitchSigmas)

Concretely: GlitchSigmas's sigmas output → SamplerCustomAdvanced's sigmas input, with RandomNoise on noise, a SamplerEulerAncestral on sampler, and a CFGGuider on guider. The pack's example API workflows show exactly this graph if you want a reference. If you find yourself wanting extra spice on the ancestral noise term, the built-in SamplerEulerAncestral already exposes s_noise - the pack deliberately dropped its own GlitchSampler node in v0.2 because that knob duplicated the built-in one. Fewer nodes, same reach.

The inputs worth knowing: scheduler (karras / exponential / normal / simple / ddim_uniform), steps, denoise (below 1.0 it does ComfyUI's standard partial-denoise tail-slicing, which is how you glitch-ify an img2img pass instead of a full txt2img), and seed for reproducibility.

Install and gotchas

Same pack, same install, and - refreshingly - no extra Python dependencies (it calls into comfy.samplers directly):

cd ComfyUI/custom_nodes
git clone https://github.com/HitmanLoges/comfyui-glitch-lab-sampler.git

Restart, look under the glitch-lab category. Verified against ComfyUI Desktop 0.22.2 with SD 1.5.

The big caveat is the same one that runs through the whole pack: this assumes diffusion-model sigma semantics. Flux, SD3, and Wan use a different parameterization, and the README says flow-model support is a roadmap item, not this release. Glitch your SD 1.5 and SDXL images, not your Flux renders. Second, the README flags that the input names of SamplerCustomAdvanced / RandomNoise / CFGGuider are the most likely point of drift across ComfyUI versions - if the node starts erroring after a ComfyUI update, that's where to look. And if you see nothing happen at all, check that jitter is above 0 before you go hunting for a bug.