DOGMA v42 HD Tile Prompt

Detail passes on hazy source material do something infuriating: they fix the haze. Locally. One tile's sky comes back crisp and contrasty while its neighbour stays soft, and the composite is a checkerboard of atmospheric conditions with glowing outlines around every object. v42 is the first composer in this pack that explicitly prompts against it.

The new input: atmosphere_lock

v42 is v36-style field extraction plus a third required string: atmosphere_lock. It's capped at 260 characters and injected as a global constraint before the protect list:

GLOBAL ATMOSPHERE LOCK: <atmosphere_lock>. Keep that atmosphere continuous across the entire tile: never locally dehaze fog/haze/smoke, never outline silhouettes, and never create bright, dark or coloured halos around objects.

That's the whole reason this node exists. Note that it's not a per-tile observation - it's global, meant to be the same string for every tile, because the failure it prevents is differential treatment between tiles. One atmosphere for the whole image, enforced tile by tile.

What else changed

The report parsing moved to a looser, more forgiving extraction. Instead of SUPPORTED: / PROTECT: it accepts VISIBLE: or SUPPORTED: for the content list and PROTECT: or AMBIGUOUS_OR_EMPTY: for the protected list, takes the last matching line rather than the first, and if it finds no VISIBLE: line at all it falls back to using the entire raw report as the visible content. Then both lists get compacted and deduped - 12 visible terms, 8 protected - via a helper that trims each term and drops duplicates.

The prompt template is also the first one here that talks about artifacts rather than objects:

Remove synthetic grain, speckling, mushy AI texture and compression-like noise while reconstructing coherent fine detail only where the source supports it.

Which is accurate about the actual problem with small-tile restoration: you get plausible mush with a grain pattern on it, not misidentified objects.

Outputs prompt and clean_report, where the clean report is compacted to VISIBLE: ... | PROTECT: ... so you can scan a run.

Inputs

tile_report, project_context (capped at 520 characters, appended last as a period lock) and atmosphere_lock. All three are required strings. Both outputs are strings.

Install

ComfyUI Manager → search DOGMA Nodes, or:

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

Restart. No dependencies - requirements.txt reads # No external dependencies. - and nothing model-side to configure. The pack's README covers only the WAN VACE prep nodes and the samplers; the semantic suite, this node included, is undocumented and registers itself from dogma_semantic_v5641.py.

Where it bites

A prompt can't fully prevent halos. If you're restoring a masked region at 4 steps against un-restored surroundings, the mask boundary is a real edge and the model will treat it as one. The atmosphere clause reduces the differential; it doesn't blend the seam. That's what the pack's edge-anchoring and statistics-lock nodes are for, and they're the ones you want if halos survive this prompt.

atmosphere_lock is easy to write badly. "hazy atmosphere" is fine; "remove the haze" is the exact instruction this node is trying to prevent, and it will be obeyed. Write it as a description of what must stay, not what to clean up.

Everything is capped, at 260 / 520 / and 12 terms. Long atmosphere descriptions get truncated mid-clause - and a half-sentence lock is a weak one.

It doesn't choose your sampler settings. Keep the Klein distilled pass at 4 steps and CFG 1; pushing CFG for "stronger adherence" makes artifacts worse, not better, on a guidance-distilled checkpoint.