DOGMA v35.1 Global Test + Full Controls

A full 8K divide-and-conquer restoration run is measured in tens of minutes. You are not going to iterate on steps, denoise and CFG at that price. This node exists because the author clearly got tired of it too: it fans one set of settings out across the whole graph, and the extra seed and center_power outputs are exactly what the pack's 2x2 tile test harness needs.

It's the controls node for the cheap iteration loop.

What it is

Structurally it's the plumbing pattern every large workflow eventually needs - one authoritative value, many consumers - rather than a Primitive node per setting. ComfyUI core gives you typed value nodes for exactly this, and the DOGMA pack gives you this: one node that emits a ready-made SAMPLER object plus four numbers, so switching a graph from "test settings" to "final settings" is editing one node.

What makes it v35.1 rather than generic is those last two outputs. seed and center_power are the inputs the tile test-lab and tile-combine nodes want - the aligned-noise node needs a seed to generate spatially consistent noise across tiles, and the combine node needs center_power to weight the blend toward each tile's centre. In the full-run workflow those values come from elsewhere; in the test harness they come from here, and having them on the same node as steps and CFG is the difference between two places to edit and six.

Inputs and outputs

Six inputs. sampler_name is the standard ComfyUI sampler dropdown (fifty entries) defaulting to euler. steps runs 1–50 with a default of 4 and no step increment, which is a wider ceiling than the v39 controls node's 20 - this one is from the era when the pack was still testing the base-model path as well as the distilled one. denoise is 0.05–1.0 at 0.01 steps, defaulting to full noise. cfg runs 0–10 at 0.1, default 1.0. seed defaults to 1976 and spans the full 64-bit range. center_power runs 1–8 at 0.25, default 3.0, and is purely a blend-shape dial for the tile combine.

Seven outputs: sampler (SAMPLER), steps (INT), denoise (FLOAT), cfg (FLOAT), seed (INT), center_power (FLOAT) and info (STRING). The info string is a formatted one-liner covering all six values, which in a 40-node test graph is worth wiring to a preview.

The SAMPLER output is the part that saves you work: it's built with comfy.samplers.sampler_object(), so it plugs straight into SamplerCustomAdvanced and any other node that takes a sampler object rather than a name.

How to use it in practice

The point of the 2x2 test lab is to run your settings on a quarter of the frame - the pack's noise and combine nodes are aligned so the test tiles reproduce what the full run would do, only smaller. So: wire this node into both the test branch and the full-run branch, tune steps/denoise/cfg against quarter-frame results, then flip the run-mode switch and let it rip without re-entering values. That's the intended loop, and it's a genuinely good one - it matches the wider lesson from tiled upscaling that the tile condition is what stops tiles diverging, and you want to know that before the 8K run.

A note on seed, since ComfyUI seeds have a reputation: the widget-level control_after_generate behaviour that randomises or increments a seed after each run applies to seed widgets, and it is the classic beginner injury in this ecosystem - you find a good seed, flip to fixed, and discover the value was already overwritten. When the seed lives on a fan-out node like this one, that trap is contained to a single node instead of scattered across the graph, which is worth something regardless of what it does downstream.

Install

ComfyUI Manager → search DOGMA Nodes (publisher axior), then restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/axior/ComfyUI-DOGMA-Nodes
# restart ComfyUI

Nothing to install alongside it - the pack declares no external dependencies and ships no weights. The checkpoints and detector models your graph loads are your own downloads.

Common issues

center_power is the output people leave dangling by mistake, because its name doesn't say tile-combine out loud. If your tile test harness produces visible seams or a hard 2x2 grid, that's the first wire to check, not the sampler settings.

The other one is version mixing. There are three controls nodes in this pack in the same family (DOGMAGlobalControlsV39, DOGMAGlobalTestControlsV351, plus DOGMAControlsV392), with different step ceilings and different strength outputs. Which one is correct depends entirely on which generation of the tile nodes you're running - v35.1's noise and combine nodes want this one, v39's detail-band nodes want the v39 controls with its mid_strength/detail_strength pair. Grabbing the newest-sounding node here is how you end up with an input that nothing accepts.