Generate Noise

Most of the time ComfyUI generates the sampling noise for you, invisibly, from the seed. Generate Noise pulls that out into a node you can see and wire. It produces a raw noise latent you control directly - for noise-injection tricks, or as the starting noise on a sampler you've told not to add its own. It's an advanced-sampling building block, not something you drop into a first workflow, and if you're not sure why you'd want hand-made noise, that's a sign you probably don't need this one yet.

It's part of KJNodes, kijai's utility pack, in the "noise" corner alongside the other low-level sampling tools.

How it works

It builds a tensor of random noise at the size, batch, and seed you specify, and hands it back as a LATENT. From there you either inject it (mix it into an existing latent to add detail or variation) or feed it as the initial noise into a custom-sampling setup where add_noise is off, so the seed and character of the noise are fully in your hands. The optional model and sigma inputs let it scale the noise to what a specific model expects rather than emitting generic Gaussian.

The inputs that matter

• width, height, batch_size, seed - the basics. Match width/height to your working resolution.
• multiplier - scales the noise strength.
• normalize - normalizes the noise; constant_batch_noise reuses the same noise across the whole batch instead of a fresh pattern per item.
• latent_channels (optional) - 4 for SD/SDXL-era latents, 16 for the newer big-transformer models (Flux, SD3, Wan-class). Get this wrong and the latent won't match your model.
• shape (optional) - BCHW for images; BCTHW / BTCHW for video latents, which carry a time dimension.
• model / sigmas (optional) - feed these to scale the noise to the model's expected sigma levels.

The output is a LATENT - inject it, or route it into your custom sampler as the noise source.

Installing it

ComfyUI Manager → search KJNodes for ComfyUI → install → restart. Or cd ComfyUI/custom_nodes && git clone https://github.com/kijai/ComfyUI-KJNodes, then pip install -r ComfyUI-KJNodes/requirements.txt, and restart.

Where people get burned

• Channel/shape mismatch. The two things that make this node error or produce garbage are the wrong latent_channels (4 vs 16) and the wrong shape (image vs video). Match both to the model you're feeding - a 16-channel Flux model won't accept a 4-channel noise latent, and a video model wants a time dimension.
• Forgetting to disable the sampler's own noise. If you're using this as the starting noise, the sampler must have add_noise off - otherwise it stacks its own noise on top and your control is meaningless.
• Reaching for it too early. This is a tool for people doing deliberate noise injection or custom sampling. For a normal generation the built-in seed handling is doing this for you already, better left alone.