DOGMA Category Settings v23 — Full-Noise Objects

Version 23 of the DOGMA settings node is the aggressive one. Where v22 massages a region, v23 tells the sampler to rebuild it from scratch at full noise - but only for the categories it deems safe to rebuild, and only inside the pixels SAM actually claimed.

The idea: full-noise, tightly bounded

The safe version of inpainting is low denoise: nudge the masked pixels, never let them drift. That's what the KB's denoise guidance is built around - 0.3-0.6 for a moderate change, and the standing rule that "the most common mistake is setting denoising too high for detail fixes." The catch is that low denoise can't do anything about structure that's wrong rather than merely soft. If a car in your plate came out fused with the one behind it, blending won't separate them. You need the model to invent new geometry.

v23 buys that by running a full-noise local reconstruction - the prompt it emits says so in caps - on a crop whose reference latent is the current master. The safety comes from two other places instead of the denoise dial: the mask, which decides the ownership boundary, and the stitch strength.

What the numbers actually say

Unlike v22, this node doesn't inherit anything. It has its own table, and the pattern is legible once you see two entries:

| family | group radius | context | target long side | crops | stitch | |---|---|---|---|---|---| | vehicles | 720 | 272 | 1984 | 6 | 1.00 | | people | 520 | 256 | 1984 | 6 | 0.98 | | street_objects | 600 | 240 | 1984 | 6 | 0.95 | | road / architecture / vegetation | 1000 | 224 | 1792 | 4 | 0.00 | | sky | 1200 | 160 | 1536 | 2 | 0.00 | | signage | 520 | 224 | 1920 | 6 | 0.00 |

Three categories come out at stitch_strength 0.00, and the emitted summary labels them SUPPORT/PROTECTION ONLY. Road, architecture, vegetation, sky and signage get detected, get masks, and get no generative local edit - the prompt literally says "ARCHITECTURE IS GLOBAL-PASS ONLY IN v23." The intended graph is: big categories get handled once, globally, at the top of the run, and the local pass only rebuilds discrete objects on top of that.

The per-category prompts are unusually specific for a node pack: roadway_people tells the model that a masked person standing in an active traffic lane who isn't on a crossing path should be removed completely and the asphalt reconstructed underneath, with no replacement person, shadow or stain. That's a real editorial policy about 1970s street photography baked into a string.

Inputs and outputs

category and restoration_brief, both strings, both normally wired from upstream planners rather than typed.

Outputs are the same eight as v22's - group_radius, context_px, max_source_side, target_long_side, max_crops, stitch_strength, edit_prompt, summary - and they go to the same places: numbers into the crop node, stitch_strength into the stitch node, edit_prompt into the local text encode, summary into your eyeballs.

Install

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

Or install DOGMA Nodes from ComfyUI Manager, or comfy node install comfyui-dogma-nodes. Zero Python dependencies - the packed requirements.txt says so and pyproject.toml agrees. No checkpoints ship with it; it configures a workflow that loads its own Klein/Flux model. MIT licensed, and built with FLUX.2 Klein 9B in mind.

Where people get burned

The 0.00 entries are the trap if you don't read them. Wire v23 into a graph that only does a local pass, ask it to fix architecture, and you'll get masks, a nice summary, and no visible change - because the node told your sampler not to regenerate it. Nothing errors. If you want the walls fixed locally, that's v24's job, or you add the global pass.

Second: context_px and target_long_side are bigger here than in v24 on purpose - v23 wants to see a lot of neighbourhood before it commits to rebuilding an object. On a card that's already tight, 1984px crops at 6 per category will make themselves known.

Third, the version pile. This pack has three settings nodes in three consecutive versions with no documentation distinguishing them, and the README covers none of them. Short version: v22 leans conservative and keeps the master as reference, v23 goes full-noise on objects only, v24 goes full-noise inside a real inpaint mask. Pick one; don't chain them.