Latent Noise Injection | Deforum

Latent Noise Injection | Deforum does exactly one thing, and the name is the whole manual: it adds Gaussian noise to a latent and hands it back. No API, no model, no hidden agenda - just samples + randn_like(samples) * noise_std. For a node that sounds almost too trivial to ship, it's genuinely useful, because animation workflows are forever fighting the opposite problem: latents that are too clean and stick to one image like glue.

You feed it a LATENT into samples and a noise_std between 0 and 1 (default 0.1). The output is the noised LATENT. At noise_std of 0 nothing happens; as you crank it up, the latent drifts further from the manifold it was sampled on, so the decoded image gets grainier and more chaotic. The honest description is "jitter": it's the quick way to make frame N+1 not be a pixel-for-pixel repeat of frame N, or to kick a stuck image-to-image pass out of a rut before the second sampler run.

Where people get burned is treating it as a substitute for real noise scheduling. Deforum-style animation isn't about adding random grain - it's about re-noising each frame by a controlled amount (the ε·σ₀ at the frame's denoise strength) so the next frame is a coherent evolution of the last, not a dice roll. That's what Prepare Latent Denoise | Deforum does, and it does it with the sampler's own sigma math. This node is the sledgehammer; that one is the scalpel. If you're building a frame loop and you want smooth motion, reach for the pair of Seed Interp Noise | Deforum plus Prepare Latent Denoise, not this. Use Latent Noise Injection when you want a quick variation pass, a one-off texture shake, or a debug step to see what a latent actually looks like with some noise in it.

It ships in deforum-comfy-nodes, so installation is the whole pack. ComfyUI Manager → search "ComfyUI-Deforum" or deforum-comfy-nodes → install and restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/deforum-art/deforum-comfy-nodes

Restart ComfyUI afterward. No model downloads, no extra Python deps - the pack is pure Python over ComfyUI's torch, and this node in particular is just a couple of tensor ops.

One honest caveat: there's no renormalization here. ComfyUI's own AddNoise node scales noise to the latent's magnitude before adding; this node doesn't bother, so larger noise_std values can push the latent visibly off-distribution and you'll see it as noise you can't denoise away. Keep it small (0.05–0.2) for anything you intend to sample again, and treat the upper half of the slider as "intentional glitch" territory. For the per-frame, sigma-aware kind of noise an animation actually wants, look at the pack's other latent nodes - this one is the quick fix, not the system.