EverAnimate Master Settings

Chunked WanAnimate workflows have a bookkeeping problem: a model here, a VAE there, a pose video, a fixed seed, four anchor latents - and you have to set all of it identically on every single chunk. EverAnimate Master Settings is the node that collects the whole run into one place and hands it downstream as a single settings socket, so your chunk nodes stop carrying twelve inputs each.

Think of it as the front of the pack. It takes the model, positive and negative conditioning, the VAE, video dimensions, the full sampler block, the anchor count, and every guide input, then bundles them into one EVERANIMATE_INITIAL_SETTINGS output. That socket feeds the Initial Chunk node, which runs chunk one. It also emits a second, lighter output - a chunks calculator socket - that drives the standalone Chunks Calculator node if you want a preview of how many chunks the run needs.

The defaults tell you what it's for

Look at the shipped values and the intended model is obvious: width 480, height 832, chunk length 81, 4 steps, CFG 1.0, LCM sampler, normal scheduler. That is a low-step distilled Wan run, the kind people actually sit around waiting for. seed is fixed at 42 by default, and the README is emphatic that you keep it fixed across all chunks - a shared seed is what keeps boundaries stable.

Two settings genuinely matter beyond the sampler block:

• chunk length - this is the per-chunk Wan window, not the total video length. Set it to 81 for a standard clip; your pose video will be split into that many 81-frame windows automatically.
• ref image background (default on) - when enabled, the background_video and character_mask inputs are ignored and the reference image's own background is carried instead. This is the single most common "why is my input doing nothing" gotcha in the pack.

What goes in

Beyond the model/conditioning/VAE trio, the optional inputs are the WanAnimate guides: reference_image, pose_video, face_video, background_video, character_mask, plus clip_vision_output and an advanced video_anchor_latent for people who want to prebuild the anchor stack themselves. pose_strength and face_strength multiply their respective guides; 0 kills the influence. num_video_anchor_latents defaults to 4, matching the EverAnimate paper's identity-anchor count.

The outputs are deliberately different types: initial settings can only plug into the Initial Chunk, and the settings socket your chunk nodes emit can't plug back into it. That's a feature, not an accident - it stops you from wiring the graph in the order that guarantees a broken run.

Install

Same as every node in the pack: ComfyUI Manager, search "Comfy EverAnimate", or

cd ComfyUI/custom_nodes
git clone https://github.com/younestft/Comfy_EverAnimate.git Comfy-EverAnimate

Restart. No dependencies, no model downloads - you still need the native WanAnimate stack (Animate 14B checkpoint, umt5_xxl CLIP, Wan 2.1 VAE) for anything downstream to actually run.

Common issues

• Forgot it feeds the Initial Chunk, not a KSampler directly. The Master never samples. It builds settings. If nothing generates, trace the chain: Master → Initial Chunk → Extension Chunks → Video Combine.
• Changing chunk length after a run changes the chunk count the calculator reports - they read the same value, so keep them in sync.
• The seed widget is marked fixed (doesn't randomize after each run) on purpose. If you're seeing different results every run, that's the thing to check first.
• This is a brand-new pack (June 2026). It's built on the newer comfy_api node API, so a stale ComfyUI can refuse to load it - update ComfyUI before filing it as broken.