⬡ Polyhedron Empty Latent

⬡ Polyhedron Empty Latent is the "one empty latent node for every family" answer to a problem you hit the moment you run more than one model architecture: SDXL wants a 2D latent, Wan wants a 5D video latent with its own frame packing, Flux2 wants a different spatial factor, and MiniMax H3 wants a joint audio/video latent. Core gives you a separate node per family. This one gives you a latent_type dropdown.

Pick Image and it emits a 2D latent for any image model - SD, SDXL, SD3, Flux, Flux2, Qwen-Image, whatever shows up later, no new node needed. Pick WAN / Hunyuan / Mochi / LTXV / Cosmos and it switches to that architecture's 5D video latent with the right frame packing, and the length control comes alive. Pick MiniMax H3 AV and you get the video+audio latent - with duration_seconds replacing the Float + Math Expression pair from the stock template (0 = off, length decides; above 0 it's clip length in seconds at 24fps, snapped up to the model's 17k+5 frame grid).

Getting the geometry exactly right

The clean way: wire your VAE in. The node probes it and the emitted latent's channel count and spatial ratio become exact for the loaded model - which also makes non-zero init noise correct (a non-16ch model with random noise but the wrong channel count is a recipe for silent garbage). Without a VAE it falls back to safe /8, 16-channel defaults, which is fine for zeros noise. Width and height snap to the VAE's spatial grid, and the node tells you what your values became via the latent_width/latent_height/width/height outputs - the latent-grid numbers (1024px at /8 = 128) are exactly what ⬡ Polyhedron Seed's preview_width/preview_height want.

The Wan I2V path

For the Wan video types this node carries full I2V wiring. Wire start_image (reference frame(s), VAE-encoded into the concat conditioning with a frame mask) and clip_vision_output (CLIP-vision features injected into both conditionings as semantic guidance), and the positive output becomes the I2V positive. That's the whole first-frame-anchored workflow in one node instead of a four-node tangle - and it matches WanImageToVideo's pin parity, so saved workflows don't break.

Init noise

noise_type writes init noise into the latent: zeros is the standard empty latent (correct at denoise 1.0 - the sampler makes all the noise), gaussian is flat white, pink/brown/blue are coloured (brown biases composition, blue biases detail), fractal is coherent multi-octave structure, and offset/pyramid round it out. noise_seed is CPU-deterministic, so the same seed reproduces across machines and PNG re-loads - set control_after_generate to fixed for a reproducible image. Effect is strongest on flow models (Wan/Flux) and at denoise < 1.0.

Installing it

Part of the ⬡ Polyhedron Suite pack. ComfyUI Manager → "Polyhedron Suite", or:

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

No extra dependencies.

Gotchas

Don't forget that length=1 on the video types gives a single-frame still - useful, but it's easy to set length on an Image-type latent where it's ignored/greyed and wonder why the video length did nothing. And non-zero init noise on a non-16ch model without a wired VAE is where you'll get weird output; wire the VAE. As with the rest of the pack, blank rendering under Nodes 2.0 means disable Modern Node Design in Settings.