⬡ Polyhedron Transition Check

Every first-to-last-frame workflow has the same silent failure mode. You gave the model two pictures, it gave you a clip, and what it actually did was dissolve one into the other over four seconds. There's no error, no warning, and if you're watching at 24 fps on a small canvas it can take a second viewing to notice. ⬡ Polyhedron Transition Check measures it instead of eyeballing it.

It works in two directions

Before a run, wire first_frame and last_frame only. It reports how much of the picture differs between the two, and whether the light differs - because little common ground between the keyframes is exactly the setup that produces a crossfade. That's a five-second sanity check that saves a five-minute render. Make the last picture from the first, as the pack's own docs keep saying, and the check agrees with you.

After a run, wire the decoded clip into images. Now it judges the passage from the clip's own first frame to its own last frame and classifies it:

• MOTION - something moved through; the in-between frames are not mixtures of the two ends. This is what you wanted.
• DISSOLVE - the in-between frames are mixes of A and B. The model faded.
• SNAP - the picture holds and then jumps at the end, reported with the second it happened. No in-between at all.

With images wired and the keyframes wired, it also tells you how close the clip's ends actually came to the keyframes you asked for - which is the difference between "the model moved between my frames" and "the model moved between something vaguely near my frames".

The one knob

pixel_level (0.08 default, 0.01–0.5) is how much a pixel has to change between first and last to count as changing, averaged over colour channels on a 0–1 scale. Higher and only the big structural changes get judged; lower and noise counts as motion. If your clip is a slow dolly in dim light, drop it. If it's a busy crowd scene, raise it or the motion score is meaningless.

Outputs

info is the verdict in words, with the measurements behind it. score is the number to watch: the median residual of the in-between frames. Near 0 means the middle frames are blends of the ends - a fade. High means they're something else entirely - a motion. -1 means it declined to judge: no clip wired, or a snap, or a still, or too short to say anything honest.

report is an IMAGE: first | middle | last frame, the pixels that change, and a small graph of progress (white) and residual (orange) over time, with the fade and motion thresholds dotted. Save it, look at it, then delete it. It's diagnostic apparatus, not a deliverable.

One structural note: this is an output node, so it executes on queue whether or not anything downstream needs it. Free, in the scheme of things, but don't wire report into a Save unless you want a PNG per render.

Why it earns its place

Ask anyone who's spent a year chaining video models and they'll tell you the same thing: the failures that hurt are the ones that look fine. Motion drift across chunk boundaries, a face quietly changing, a run that technically satisfies the prompt by crossfading. All of those are measurable with a residual curve and invisible in a contact sheet.

ControlNet gets a strength dial and start/end fractions so you can ask for 40% of a pose. Keyframe workflows have had nothing equivalent - the model either moved or it didn't, and you found out by watching. A number that says "this was a dissolve" belongs in the graph next to the render.

MiniMax Keyframes → Sampler → VAE Decode → Transition Check
       │  first_frame/last_frame ───────────────┘
Transition Check.report → Save (when you want the picture)

Install

Manager, search "Polyhedron Suite", or:

cd ComfyUI/custom_nodes
git clone https://github.com/PolyhedronAI/ComfyUI-PolyhedronLoRAStack.git
# restart ComfyUI

No dependencies required - the pack's requirements.txt is a comment block saying as much. No model file either: this is arithmetic on tensors you already have, which is why it's the cheapest node in the pack to try.

Where it goes wrong

Judging a clip with a cut in it. A two-shot clip with a hard edit in the middle reads as a snap or a nonsense motion score. One shot per check.

Very short clips. Four frames isn't a passage; you get -1 and a shrug, correctly.

Interpreting score as quality. It isn't. A high score means "not a fade". Whether the motion is good is still your eyes' job - this just stops you shipping a dissolve you didn't ask for.