DOGMA v43 Base Local Prompt

If you're picking one of this pack's six local-prompt nodes and you don't want to think about it, pick this one. DOGMALocalPromptV43 throws away your denoise knob and uses a fixed table instead. That sounds like a downgrade and is actually the feature.

What changed from v42

v42 took your base_denoise and multiplied it by a per-category factor - you tuned one number, the node scaled it. v43 keeps base_denoise as a required input (so it wires the same way, and it's still the fallback for categories the table doesn't know), but ignores it for anything in the table. sky is 0.12. water 0.14. road 0.16. grass 0.20. trees 0.22. buildings 0.27. clothing and food 0.28. people, animals, products 0.30. furniture and machinery 0.32. cars, bus and trucks 0.35.

Those are the values the author settled on, and the info output tells you the configuration they were settled for: Distilled9B CFG1. In other words, the whole table assumes a distilled FLUX.2 Klein 9B running at CFG 1 - the low-step, no-guidance setup. That's the honest read: this is a working configuration someone tuned by running it, not a general theory of denoise.

The prompts changed too. Banked objects now get an explicit "do not add, remove or move an instance", while sky, vegetation, road/water and architecture get their own dedicated bodies: sky talks about colour, exposure, haze, cloud structure and softness with no new clouds or light sources; architecture talks about facade geometry, window count and identity; road and water are told to preserve exact geometry, markings and reflections. The headers stay blunt - 1970s ITALIAN CARS. and friends.

Outputs

Three this time, and the third is the one to wire: prompt, info, denoise. The denoise FLOAT is already category-correct, so send it straight to your local pass's strength input. If you want to tune the run, tune the prompt or the mask - not this.

Why fixed values are defensible

Because a whole-scene pass at one denoise is a category error, and a tunable per-category scale is a knob you'll get wrong. With a multiplier you have to reason about both the base and each factor. With fixed values you have a table someone already ran against the model you're using - distilled Klein at CFG 1 - and the numbers match the conventional wisdom for masked re-rendering: low enough that the source survives, high enough to rebuild structure where structure exists.

Note the v43 family is self-consistent: this node, DOGMASceneInventoryPlanV43 (which emits the categories, kinds and thresholds) and DOGMARegionStitchV43 (whose feathering is kind-aware - a hard 0.96 blend for objects, 0.72 for surfaces) all share the same vocabulary and the same assumptions. Mixing v43 prompts with a v40-era mask and a v35 stitch is how you get results that look almost right.

Inputs and outputs

category, kind, project_context, base_denoise in - all STRING or FLOAT inputs you wire from the planner. prompt, info, denoise out.

Install

ComfyUI Manager → search DOGMA Nodes, or:

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

Restart ComfyUI. Zero pip dependencies, zero model downloads - the pack is glue over PyTorch and ComfyUI's sampler APIs. The Klein checkpoint and a masker (SAM or equivalent) come from elsewhere in your graph.

Where it bites

Run it with the base Klein model instead of the distilled one and 0.35 on a car is noticeably hotter than the table assumes; you'll see the panel lines move. That's a model mismatch, not a node bug.

The other trap is kind. OBJECT, STRUCTURE, SURFACE - as a string. The scene planner emits it per slot, so wire it. Typing object lowercase costs you the special-cased bodies and you'll blame the denoise.