VOID PQ5 Settings

This is the pack's settings-bundle node, and it exists purely for graph ergonomics. The VOID PQ5 sampler already exposes num_inference_steps, guidance_scale, strength, and seed directly on its own inputs - but if you're doing the two-pass VOID workflow, you don't want to edit the same four numbers in two places every time you iterate. This node packages those four values into a single settings object you can feed into the sampler's optional settings input.

Nothing is computed here. No models, no tensors, no GPU. It's a struct in node form. Think of it like a primitive-input node you can route around the graph - you can even use it to keep your "one settings bundle" pattern consistent if you later build bigger graphs around the sampler.

The inputs that matter

• num_inference_steps (INT, default 50) - sampling steps. VOID's default is 50; fewer is faster and rougher, more is slower and cleaner. This is a video model, so step count multiplies against a heavy 3D transformer - don't casually double it.
• guidance_scale (FLOAT, default 1.0) - CFG. Note the default is 1.0, not the 7-ish you're used to from SD. The VOID pipeline treats guidance > 1.0 as "enable classifier-free guidance" (the prompt encoder checks this), and this model family is tuned around low or zero CFG.
• strength (FLOAT, default 1.0) - how much of the original video's latents to keep. 1.0 = full regeneration within the mask; lower values blend toward the input, which is useful for pass 2 when you want the refinement to stay close to pass 1's output.
• seed (INT, default 42) - reproducibility.

Output: a single settings wire (PQ5_SETTINGS).

How the override works

When you connect settings into the sampler, it overrides the sampler's own direct inputs - the sampler reads the bundle and replaces its step/guidance/strength/seed values with the bundle's. So you can keep the sampler node's four inputs untouched (or disconnected) and drive everything from one settings node. If you leave settings unconnected, the sampler just uses its own inputs. Two ways to do the same thing; pick one per graph and don't mix, or you'll forget which is winning.

Install & notes

Same as every node in the pack: ComfyUI Manager → search "ComfyUI-NetflixVoid", or git clone https://github.com/jaskirat05/ComfyUI-NetflixVoid into custom_nodes, then restart. It's a pure struct node - instant to run, no dependencies beyond the pack itself.

One practical tip: because guidance_scale defaults to 1.0, a settings bundle you build for pass 1 will happily carry that low guidance into pass 2. That's usually what you want (the model family is tuned for it), but if pass-2 output looks washed out, check which value is actually reaching the sampler - it's easy to forget the settings bundle is the one in charge.