MetaView Sigmas

Most of the time you don't think about sigmas - you pick a scheduler by name in KSampler and move on. MetaView Sigmas (class MetaViewSigmas) exists because the MetaView pipeline is not most of the time. Its rendering model was distilled down to a specific 8-step schedule, and if you hand it the wrong schedule you get a worse image or a broken one. This node emits the exact Qwen-Image / MetaView FlowMatch sigma schedule the model was fused for, and it's deliberately small: one output, a few inputs, no ceremony.

How it works

It's a pack-local SIGMAS node - it doesn't register into ComfyUI's global scheduler table, which is a deliberate design choice. KSampler's named scheduler dropdown can't represent this schedule, so the pack emits it as data you feed to a sampler that accepts a raw SIGMAS input. The schedule is an exponential mu computed from the patch-grid length (the generation resolution divided by 16, squared), with a terminal shift to 0.02.

The wrinkle is that mu depends on your actual image grid. That's why latent is the preferred input: the node reads the latent's spatial size, converts back to pixels (×8), and derives mu from the real generation grid so the schedule matches what the model is actually denoising. The width/height widgets are just fallbacks for when you can't wire a latent.

Inputs and output

• steps - default 8. That's not a coincidence; the MetaView weights are fused for 8 steps. You can crank it, but you're leaving the tested regime.
• latent - optional, but preferred. Feed it the latent output from MetaView3DConditioning so mu is derived from the real grid.
• width / height - fallbacks (defaults 960×528) when no latent is wired.
• sigmas - the output. It goes into SamplerCustom or SamplerCustomAdvanced, not KSampler.

Wiring

The MetaView sampler recipe is fixed by the model: 8 steps, CFG 1.0, euler. The sigmas output feeds SamplerCustom's sigma input, and the latent_image for that sampler must come from MetaView3DConditioning's latent output so everything agrees on grid size.

Installing and troubleshooting

Same pack as the rest of MetaView: ComfyUI Manager → "Enviral Design Node Pack", or clone into ComfyUI/custom_nodes and restart. This node needs no models of its own - it's pure schedule math.

The classic mistake is trying to force this into KSampler. Don't. KSampler's scheduler dropdown and this SIGMAS node are two different mechanisms; this one plugs into SamplerCustom/SamplerCustomAdvanced. And don't swap in an EmptyLatentImage where the conditioning node's latent should go - if the sigma mu is computed from the wrong dimensions, you'll get a schedule tuned for a grid you're not sampling.

If you're new to samplers, think of it as: the MetaView model is a picky actor with a set rehearsal - 8 steps, euler, this schedule, this latent. Give it its script and it behaves. Improvise and it gets stage fright.