Sample Settings 🎭🅐🅓

Displayed as Sample Settings, this is the node the README points to when it talks about "custom noise scheduling via Noise Types, Noise Layers, and seed_override/seed_offset/batch_offset." It's the advanced-tuning bundle, not something you need for a basic AnimateDiff run - plug it into a loader's or Use Evolved Sampling's sample_settings input once you've got a workflow producing motion and want to fix a specific problem with how noise or seeding behaves across frames.

The headline feature buried in here is FreeNoise: per the README, "FreeInit and FreeNoise support (FreeInit is under iteration opts, FreeNoise is in SampleSettings' noise_type dropdown)." If you've read about FreeNoise elsewhere as a way to get better temporal consistency without retraining anything, this node's noise_type field is where you turn it on.

How it works

Standard diffusion samples fresh Gaussian noise per step. Across a video batch, how that noise is generated, offset, and correlated between frames has a real effect on flicker and consistency - this node exposes the controls for that, plus a few adjacent knobs (CFG scheduling, sigma schedule, mid-sampling image injection) that are all really "advanced sampling behavior" rather than anything specific to the motion module itself. That's consistent with Use Evolved Sampling supporting this even outside AnimateDiff entirely.

The inputs and outputs that matter

Required, and worth understanding even at a glance: noise_type (a dropdown - this is where FreeNoise lives), seed_gen (how per-frame seeds are derived), batch_offset and seed_offset (integers for shifting which noise seeds are used, useful if you want a different but reproducible variation of the same batch).

Optional, and each one is its own rabbit hole with a dedicated feeder node: noise_layers, iteration_opts (FreeInit lives here), seed_override, adapt_denoise_steps (boolean), custom_cfg, sigma_schedule, image_inject (mid-sampling image injection, per the README's feature list), and ancestral_opts. Output is settings (type SAMPLE_SETTINGS), which only fits a loader's or Use Evolved Sampling's sample_settings slot.

How to install it

ComfyUI Manager: search AnimateDiff Evolved by Kosinkadink, install, restart. Manual:

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

Nothing extra to download for this node - it's pure sampling configuration layered on top of whatever motion setup (or plain model) you've already got working.

Common issues & troubleshooting

This feels overwhelming for a first AnimateDiff workflow. That's a fair read - it is. Get a basic loader working first with defaults everywhere, confirm you're getting coherent motion at all, and only reach for this node once you have a specific complaint (flicker, seed variance you don't want, wanting FreeNoise specifically) to fix.

Not sure if a change here actually did anything. Because most of these inputs take dedicated feeder nodes rather than raw values, it's easy to leave one unplugged and assume you configured it. Double-check each optional input you care about is actually wired to something, not just present as an empty slot on the node.

We don't have documented community failure modes specific to this node beyond general confusion about its scope - it's advanced enough that most public workflows don't touch it, which is itself useful information: you probably don't need it for a first pass.