LatentNoisy (🚫Deprecated)

The pack's own README calls this out directly as one of two ways to inject noise into a latent manually - the other being easy preSamplingNoiseIn - and its changelog is just as direct about which one to actually use: easy preSamplingNoiseIn was added specifically to "replace the easy latentNoisy node that needs to be fronted to achieve better noise injection." That's about as clear a deprecation note as a changelog gets. This node still works, but the pack itself is telling you there's a better version.

What it did

Rather than running a full sampling pass, easy latentNoisy computes a noisy latent directly at a specific point in a chosen sampler/scheduler's noise schedule - useful for workflows that want to seed a partial-denoise step (like a hires-fix pass) with a latent that's already at the "right" noise level for where sampling should resume, instead of starting from pure noise or from a fully clean image.

The inputs and outputs that matter

• sampler_name / scheduler - which noise schedule to compute against. A long list of samplers (44 choices) and 9 schedulers, matching what you'd see on a standard sampler node.
• steps - the resolution of the schedule you're slicing into (note the default sits at 10,000, far higher than a typical per-run step count - treat this as defining how finely the underlying schedule is subdivided, not as "how many steps this generation runs").
• start_at_step / end_at_step - where within that schedule to land.
• source - CPU or GPU, where the noise generation happens.
• seed - controls the noise pattern.
• Optional pipe, optional_model, optional_latent - feed in an existing pipe/model/latent to noise, rather than starting from scratch.

Outputs: pipe (pass-through), latent (the resulting noisy latent), and sigma (FLOAT - the noise level at that point in the schedule, which some downstream custom-sampling setups need explicitly).

Why it's deprecated, and what to use instead

Per the pack's own changelog, easy preSamplingNoiseIn does the same conceptual job - injecting noise into the latent space - with a design the author explicitly called better, and it's meant to sit in the normal presampling position in your graph (between loader and sampler) rather than needing to be "fronted" the way this node did. If you're building fresh, start with easy preSamplingNoiseIn and only reach for easy latentNoisy if you're maintaining an older workflow that already depends on its specific sigma output.

Installing it

Ships with the base pack. ComfyUI Manager: search ComfyUI Easy Use, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/yolain/ComfyUI-Easy-Use

then install.bat on Windows or pip install -r requirements.txt, restart. No models - this is schedule math, not a model-dependent operation beyond needing to match whatever model you'll actually sample with.

Common issues & troubleshooting

Confused by the huge default steps value. 10,000 is not a real per-run step count - it's the granularity of the schedule this node is computing against. If your images look wrong after using this node, check start_at_step / end_at_step relative to that scale, not relative to the 20-30 steps you'd normally set on a sampler.

Sampler/scheduler mismatch with the actual sampling step downstream. The noise level this node computes is only meaningful if the sampler that consumes the resulting latent uses a matching sampler/scheduler configuration. If you set one schedule here and sample with a different sampler or scheduler downstream, the noise level won't line up with what that sampler expects at its starting point, and results will look off in a way that's hard to diagnose without checking this mismatch first.

Considering this node for a new workflow. Don't, per the pack's own guidance - go straight to easy preSamplingNoiseIn unless you specifically need the sigma output this node exposes and can't get it another way.