Noise Layer [Add Weighted] 🎭🅐🅓

Noise Layers are how the README's "custom noise scheduling via Noise Types, Noise Layers, and seed_override/seed_offset/batch_offset" actually gets built - stack them together, each with its own noise type, seed behavior, and optional mask, and you're composing a custom noise recipe for your sampling instead of accepting a single flat draw. This variant blends its noise into whatever layers came before it at an explicit weight you control, then gives you a separate dial to rebalance the total afterward.

It's one of three closely related noise-layer nodes in the pack - this one, Noise Layer [Normalized Sum], and Noise Layer [Replace] - and picking the right one matters more than any individual setting on it.

How it compares to its siblings

Replace fully overrides the previous layer within its region - no blending. Normalized Sum blends layers together with weights that are automatically kept in proportion, so the combined result stays balanced without you managing it. This one, Add Weighted, sits between those: you set an open-ended noise_weight (0 to 10, not capped at 1 like Normalized Sum), and because stacking weighted noise this way can push the combined result's overall variance out of whack, balance_multiplier gives you a separate dial to rescale the total back down afterward. It's the more manual, more flexible option of the three - reach for it when you specifically want a layer to dominate or barely register, rather than a clean proportional mix.

The inputs and outputs that matter

Required: batch_offset (default 0) - shifts which frame this layer's pattern starts from, noise_type - the noise-generation strategy for this layer, seed_gen_override - how this layer derives its seed independent of the base sampling seed, seed_offset (default 0), noise_weight (default 0.5, range 0–10) - how strongly this layer blends in, and balance_multiplier (default 1) - rescales the combined result after blending.

Optional: prev_noise_layers (NOISE_LAYERS) to chain another layer underneath this one, mask_optional (MASK) to restrict this layer to a region, and seed_override (default 0) for a fully explicit seed rather than an offset.

Output is a single NOISE_LAYERS, which only fits Sample Settings' noise_layers input.

How to install it

ComfyUI Manager: search AnimateDiff Evolved, confirm the author is Kosinkadink, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/Kosinkadink/ComfyUI-AnimateDiff-Evolved

then restart. No model download tied to this node - pure noise configuration on top of whatever motion setup you already have working.

Common issues & troubleshooting

Added a layer and nothing visibly changed. Confirm the NOISE_LAYERS output reaches Sample Settings' noise_layers slot, and that Sample Settings itself feeds your loader or Use Evolved Sampling. It's an easy chain to leave half-connected, since none of these error loudly when unplugged.

Result got noisier or flickerier after stacking layers. That's noise_weight pushed too high relative to balance_multiplier - try lowering the weight, or raising the balance multiplier to bring the combined variance back down.

Not sure whether this is the noise-layer node you actually want. If you want a clean, automatically-balanced blend, use Normalized Sum instead - it removes the manual rebalancing step this node requires. If you want one layer to fully override another rather than blend, use Replace.