LatentMultiply

Latent Multiply scales every value in a latent by a single multiplier - the most direct possible way to change the "energy" of the signal the sampler is about to work with. It's the latent-space equivalent of turning brightness or contrast up and down, except it's not really either of those things. It's multiplication, applied to the whole tensor, and that makes it a useful but dangerous toy.

The inputs are just samples and multiplier - a float from -10 to 10, default 1.0 (which is the no-op). Output is the scaled LATENT. Set it to 0 and you get a blank latent; negative values invert the signal.

What it's actually for

Three honest uses. Adjusting img2img strength: multiplying the source latent before a second sampling pass nudges how strongly the original content asserts itself through denoising - a 0.9–1.1 tweak can move the result without touching the denoise slider. Latent-space style math: combined with Subtract and Add, scaling is how you control how much of a computed "direction" you apply - half the direction is multiplier 0.5, not a full application. And occasionally, rescuing an overexposed or underexposed generation: if a model has a known bias toward washed-out output, a small pre-decode multiplication is a cheap correction. Note that last one is genuinely rare - most "washed out" problems are VAE issues and multiplying the latent will not fix those.

The trap

Latents are trained distributions, and this node kicks the tensor outside it. The multiplier range goes to ±10, but anything past about 1.2–1.5 starts producing artifacts the model never learned - blown-out colors, posterization, or noise - because you're amplifying the noise component of the latent right along with the signal. There's no clean mapping from "multiply by 1.3" to "a bit brighter"; it's a physical thing, not a perceptual one. So the honest advice: treat 0.8–1.2 as the usable band, and if your intent is a genuinely different brightness, do it in pixel space after decode or change the model's conditioning instead. This node is for people who know they want the latent louder, not for casual brightness tweaking.

Also worth noting: it operates per-channel uniformly, so it can shift color balance in unpredictable ways if the model's channels carry different semantic weight. And it's batch-agnostic - multiplier applies to every item in a batch the same way, which is what you usually want.

It ships with ComfyUI core, no install. Use it sparingly, keep the multiplier near 1, and pair it with Latent Subtract and Latent Add when you're doing real latent arithmetic - it's the volume knob of that toolbox.