DOGMA Category Settings v13

Version 13 keeps v12's interface byte for byte - one category input, nine outputs ending in stitch_strength - and changes only the numbers. That sounds like a nothing release. It's actually the most opinionated one in the set, because of a single comment left in the table: "Architecture was too close to the source in v12. Still masked and identity-preserving, but now allowed to visibly fix defects."

What got retuned

Everything, gently, in one direction: more reach, more region, more crops. Compare a few rows between v12 and v13:

• vehicles - threshold 0.14 → 0.13, group radius 320 → 360, source bound 2400 → 2700, max crops 8 → 10, alpha 1.00 → 1.00
• roadway_people - 0.18 → 0.16, radius 260 → 340, alpha 1.00 → 1.00
• architecture - threshold 0.34 → 0.28, radius 720 → 820, context 128 → 160, source 2800 → 3200, model side 1344 → 1536, crops 6 → 8, alpha 0.55 → 0.82
• road - 0.34 → 0.32, radius 800 → 900, source 2800 → 3200, model side 1152 → 1280, alpha 0.40 → 0.35
• sky - radius 900 → 960, context 64 → 96, model side 1024 → 1024, alpha 0.35 → 0.35
• water - context 96 → 112, model side 1152 → 1280, alpha 0.50 → 0.50

Denoise stays 1.00 across the board (the v12 doctrine holds), and summary still spells out "Klein full 4-step denoise=1.00, final alpha=X".

The architecture row is the headline. In v12, facades were pulled back so hard - alpha 0.55, model side 1344, threshold 0.34 - that a genuinely malformed building came out barely touched. v13 raises the blend to 0.82 and the model side to 1536, which is the difference between "preserved" and "repaired". Meanwhile road and sky got slightly more conservative relative to their own reach (road's alpha drops to 0.35), because those are the surfaces you can scorch with too much generated texture.

Read together, v13 is the author's settled position on the trade: objects and structures should visibly improve; huge flat fields should not.

Wiring, and choosing between the versions

Same fan-out as v12: sam_threshold to the segmenter, group_radius/context_px/max_source_side/target_long_side/max_crops to the cropper, stitch_strength to the blend. The extra crops (10 instead of 8) and larger radii mean more sampling passes per run - this version costs more, and it costs the most exactly on the scenes it helps most, i.e. cluttered ones.

Which should you pick? If you're on a modern edit model and care about buildings, v13. If your runs are already long and you want the cheapest reasonable result, v12. The pack ships every version deliberately, in the same menu, and old workflows keep loading because the author patches older planner classes in place rather than renaming them - so picking a version is a real choice, not a leftover.

Install

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

Or Manager → DOGMA Nodes, restart. Nothing to install beyond that; the pack has no Python dependencies at all, which is pleasantly unusual. Model requirements live outside the pack: a VLM for planning, a grounded segmenter for masks, and a Flux-family edit model for the passes.

Common issues

Buildings still look untouched. Then your crop isn't reaching them. max_source_side at 3200 is advisory for the cropper - if the region is downscaled well below target_long_side before the pass, there's nothing left to fix. Check what the crop node reports.

Sky banding. Alpha 0.35 still means a third of the pixels came from a fresh generation over a smooth gradient. If you see bands, drop that row's alpha or exclude sky entirely - the none row exists for a reason.

Slower runs than v12. Ten crops per sector instead of eight, larger crops, higher model side. Expected, and it's the cost of the retune.