Flake Animate Preset

ComfyUI Flakes is built on one idea: the stuff that defines what you're generating lives in small YAML files instead of being wired into the graph. Flake Animate Preset is where the other half of that lives for video - the machine settings. Which diffusion model(s), which text encoder, which VAE, how many steps, what sampler, how long the clip runs, what resolution. It doesn't sample anything. It just reads a preset off disk, loads the files with ComfyUI's core loaders, and hands you a flake_anim_data pin that the rest of the animate chain passes along.

Why bother? Because image-to-video is the place where you least want to rewire a graph. Swapping from a Wan 2.2 two-model setup to a MiniMax H3 single-model setup normally means rebuilding half the canvas. Here it's a dropdown.

The two families it knows

The family widget is the switch, and it decides which files the node expects:

• Wan 2.2 (wan22) wants a high-noise and a low-noise diffusion model, a Wan text encoder, and the Wan VAE. That's not redundancy - 2.2 is a two-pass architecture where the high-noise expert handles motion and composition and the low-noise expert refines detail, so the preset carries both files and a refiner_step saying where the handoff happens.
• MiniMax H3 (mh3) wants one diffusion model, a MINIMAX-type text encoder, a video VAE and a separate audio VAE. H3 generates stereo audio as part of the same pass, which is why there are two VAEs to point at.

Numerics you leave out of the YAML fall back to family defaults: 4 steps at CFG 1 with euler/linear_quadratic at 16 fps and 5 seconds for Wan 2.2, 4 steps at CFG 1 with euler/simple at 24 fps and 5 seconds for H3. Those are distilled-model settings - don't "helpfully" raise steps to 30 and expect free quality.

Inputs and outputs

Three fields, and only one of them is really a choice:

• family - Wan 2.2 or MiniMax H3.
• preset - the dropdown of animate presets found under ComfyUI/models/flakes/model_presets/video/<family>/. You build them in the pack's own preset editor, not by hand-editing YAML until you know the format.
• overrides_json - optional, and the author's tooltip calls it: "JSON override of steps/cfg/sampler/scheduler/filename_prefix. Managed by the Flake widget." You don't type this. It's how the grid's per-item tweaks get stored.

One output: flake_anim_data. Wire it into Flake Animate Stack or Flake Animate Combo if you want character/style flakes layered on, or straight into Flake Generate Animation if you just want to render the preset as-is.

Install

Via ComfyUI Manager, search ComfyUI Flakes, or:

cd ComfyUI/custom_nodes
git clone https://github.com/JeyzerMC/comfyui-flakes

Restart ComfyUI. requirements.txt is one line - pyyaml - which already ships with ComfyUI, so there's no pip step to get wrong. Weights go in the usual places (models/diffusion_models, models/text_encoders, models/vae); the preset stores file names, not paths, and resolves them against those folders.

One version note before you start: the animate nodes lean on core ComfyUI nodes that only exist in recent builds (Wan's image-to-video conditioning, the MiniMax H3 nodes). On an older install the failure is a missing-node error, not a friendly message.

Where people get burned

The obvious one: if the preset dropdown says Select a preset... or No model preset is selected, that's the node telling you it found nothing on disk. Selecting the placeholder and hitting Queue raises No animate preset is selected - pick one from the dropdown. Make a preset first.

A preset with a missing model raises by name - Animate preset '<x>' has no 'diffusion_model_low' set - pick one in the preset editor - which is about as good an error as you'll get in this ecosystem.

Two quieter ones. First, animate LoRAs patched by the preset hit the diffusion model only; the text encoder isn't patched, so trigger words still have to be in the prompt. Second, and this is not Flakes' fault: MiniMax H3's community licence excludes the US, EU, UK and South Korea from its grant, outputs included. The weights run fine anywhere; the licence is a different question, and it's worth knowing before you build a pipeline around that family.

Last thing, and it's the nice one: this node re-runs when the preset's YAML changes on disk. Edit a preset, queue again, and flake_anim_data reflects the edit without a restart.