Noise Control Script

Ever tried to reproduce an Automatic1111 image in ComfyUI and gotten something close but wrong? A big chunk of that gap is where and how the initial noise gets generated. Noise Control Script gives you control over the sampler's RNG source and a couple of A1111-compatibility tricks, so you can actually match results across the two. It's a Script node in the Efficiency Nodes pack, meaning it modifies how an attached Efficient KSampler behaves rather than doing anything on its own.

How it works

Its SCRIPT output plugs into the script input of an Efficient KSampler, and from there it hijacks a few things about how noise is made. Because it's a script node it can be chained with other scripts (Tiled Upscaler, XY Plot) on the same sampler.

The headline feature is RNG parity with A1111. Comfy and A1111 historically differed in whether initial noise is generated on CPU or GPU, which alone can change your image. This node lets you pick.

The inputs and outputs

• rng_source (cpu, gpu, nv) - where the noise comes from. For A1111 matching, gpu. nv is NVIDIA's generator.
• cfg_denoiser (default false) - the smZ-developed CFG Denoiser hijack, which gets you closer to A1111's actual sampling math. Big caveat straight from the README: it does not work with several conditioning types, including ControlNet and GLIGEN. Turn it on for a plain txt2img A1111 recreation, leave it off the moment you add ControlNet.
• add_seed_noise (default false) plus seed and weight (default 0.015) - this is the cg-noise "seed variation" trick: blend a second seed's noise in at a small weight to get controlled variations on an image you already like. Keep weight low; 0.015 is a nudge, not a reroll.

The output is SCRIPT, and it only does anything once it reaches an Efficient KSampler.

The README even hands you the full A1111 recipe: encode with length+mean token normalization and A1111 weight interpretation (via an Efficient Loader), set rng_source to gpu, and turn cfg_denoiser on.

Installing it

ComfyUI Manager → search Efficiency Nodes for ComfyUI, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/jags111/efficiency-nodes-comfyui

Restart ComfyUI afterward. No extra downloads for this node.

Common issues

The single biggest gotcha is the cfg_denoiser incompatibility. If you've got it on and you're using ControlNet or GLIGEN, expect broken or ignored conditioning - that's documented behavior, not a bug. It's strictly a plain-conditioning tool.

Second, seed variation confusion: add_seed_noise does nothing unless it's toggled on, and even then a weight that's too high stops being a "variation" and just becomes a different image. Start at the 0.015 default and creep up.

And remember it's a script node - nothing happens unless the SCRIPT output feeds an Efficient KSampler. A stock KSampler has no script input.

Pack-wide: IMPORT FAILED at startup (the recurring pip freeze non-zero exit) drops the entire pack. Update ComfyUI, update the node to the latest commit, and confirm your Python environment is healthy.