Flake Animate Combo

Flake Animate Combo is Flake Animate Stack with one word changed. In the shipped source it's class FlakeAnimCombo(FlakeAnimStack): pass - the Python is identical, and the whole difference is in how the browser queues it, which turns out to be a genuinely different tool.

Stack versus Combo, in one sentence

A Stack runs every active flake together; a Combo runs them one at a time, once each, as separate jobs. Same grid, same YAML flakes, same flake_anim_data output. The frontend keeps a single active flake in flakes_json per queued combination, so you get the cross-product instead of the combination.

That distinction is the whole reason video needs both. On the image side, "all at once" is usually what you want - the flake is a character plus an outfit plus a style and they should all apply. On the animate side you're frequently doing the opposite: same start frame, five different characters, or one character across six poses, and you want five or six separate clips out the other end, each labelled.

What it does with your images

Each grid item carries its own start images, and each image becomes its own job. So the arithmetic is flakes × images: three active flakes with two images apiece is six clips, not two. Results stay keyed to their grid item, so if you drag items around or duplicate one, the clip that's already rendered doesn't jump to a different item.

The comparison that makes this worth the VRAM: drop in one strong pose image, add four of your character flakes, queue once, and you're comparing four characters on identical motion. Or fix the character and stack four different start frames and you're comparing shots. It's the "same seed, different subject" test that everyone runs by hand on the image side, batched.

Inputs, outputs, install

Same as the stack: family (hidden and inferred from the connected preset), flake_anim_data, flakes_json ("JSON-encoded list of animate flake entries. Managed by the Flake widget"), one output flake_anim_data into Flake Generate Animation.

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

Restart afterwards, or install via Manager by searching ComfyUI Flakes. No pip dependencies - the pack's requirements.txt is pyyaml, which ComfyUI already has.

The traps, which are all about queue size

Every active flake in the grid multiplies the run. Bypass the ones you're not testing, and read the queue count before you press the button - the failure mode isn't an error, it's a GPU that's busy for two hours with a combination you didn't want.

Everything from the stack still applies: LoRA keyframe schedules are ignored on the animate path (LoRAs apply at plain strength), the text encoder isn't patched so keep trigger words in the prompt, and a missing start image surfaces as a complaint from Generate Animation rather than a friendly warning here.

And one design note worth respecting: this node doesn't rewire anything for you. It queues the same graph repeatedly with a different flake selected. That's why it's cheap to add and why an "All Variants"-style sweep still costs one full video generation per combination.