⬡ Polyhedron Sampler

The Polyhedron Sampler is the pack's one-node answer to "which sampler do I use?", and the honest headline is that in Single mode it's a superset of KSampler plus KSampler (Advanced) - all the fields you know, plus the ones you always wished the basic node had. But the reason it exists is the High + Low pill, because that's the architecture Wan 2.2 ships and most of the ecosystem still struggles to wire it by hand.

Single mode

seed, steps, cfg, sampler_name, scheduler, denoise - all there. What the stock Advanced node doesn't expose is a real denoise field alongside start_at_step / end_at_step and return_with_leftover_noise, and add_noise to decide whether the run adds noise at all. That combination turns the manual HIGH/LOW split (two chained samplers, one handing an unfinished latent to the next) from a multi-node dance into one node's fields. sigma_shift applies a flow-matching shift before the schedule is built - 0 means off, and the tooltip is blunt: Wan 2.2 needs a shift, typically 8.0, or a short run leaves residual noise as RGB speckle. This is the same shift knob the flow-matching world learned to care about when scheduler choice stopped being the tuning lever.

High + Low mode

Flip dual_moe on and the same node runs a noise-split two-expert denoise. model becomes the high-noise expert, model_low takes the low one, and the boundary (the "Handoff") is the one number that matters. It's a noise level, not a step count: every step whose sigma is at or above it runs on the high expert, the rest on the low. There's no automatic value - the Wan 2.2 conventions are 0.875 for text-to-video and 0.900 for image-to-video, so start there and adjust by eye.

cfg_low gives the low expert its own guidance - distilled and Lightning LoRAs typically run the low pass at 1.0, which is exactly the "speed LoRA on the low pass only" pattern the community converged on. sampler_low and scheduler_low default to "same as high". handoff_mode decides the seam: Continuous runs both experts as one unbroken denoise, while Wan MoE parity reproduces the Wan MoE KSampler exactly (high expert finishes on a clean estimate, low re-noises it with the same seed noise) - use that one when you need to match Wan 2.2 res_2s output byte-for-byte.

You can also drive it with external schedules: sigmas (one full curve, split at the handoff), or sigmas_high/sigmas_low as separate segments, or a custom noise source. Output is a single LATENT, image or video, whichever you fed in.

Live preview

preview_mode picks the in-node decoder: Still · refreshes one frame per step like a KSampler; the Video · modes animate every frame of the predicted latent so on a video you watch the motion settle before the run ends. The two Video · TAE modes decode through a small approximate decoder - slower per step, far closer to the finished picture, and they carry their download size in the label because the files have to be on disk first. It's visualisation only - the output latent is identical whatever you pick.

Installing it

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

Or ComfyUI Manager, search "Polyhedron Suite". No hard dependencies. The recurring gotcha is the shift: if you're coming from a workflow that used a separate ModelSamplingSD3 node, leave sigma_shift at 0 (an upstream node already provides it) - a positive value here overrides rather than adds. And when driving with an external SIGMAS curve, leave the shift at 0 too; the curve owns the schedule.