DOGMA AuditBatch v1.0.14

The bug this version exists to fix

Every mask-audit prompt in this family used to carry an implicit assumption: the target is a thing, and background classes - ground, walls, sky - are what a bad mask accidentally grabs. That assumption falls apart the moment you're segmenting a street surface or a facade on purpose. A "people" mask that selects pavement is wrong. A "road" mask that selects pavement is correct. Same pixels, opposite verdict.

DOGMAAuditBatchV114 is the category-relative rewrite. Its module docstring says it in one line: ground/architecture are valid when they are the target. If you're auditing anything other than objects, this is the one to reach for.

What changes in the prompt

The node takes the category string and prepends targeted guidance before assembling the instructions:

• For road/paving/ground/asphalt/cobble/floor/soil/street: the target is a continuous ground surface, not a bounded object. Darkness, wet reflections, paving joints and shadows don't change the category.
• For building/architecture/arch/facade/colonnade/cathedral: walls, roofs, columns and facade components belong to it; real openings, occlusion and partial structures are valid.

Then the general rules: judge the selected pixels, not everything visible in the crop; PASS when substantial selected regions belong to the target; FAIL only if substantial selected regions clearly belong to a different category or no target is visible; name the wrongly selected category in the reason. And a line that only exists here: never reject the target itself merely because it is ground, architecture, sky or background.

A third verdict is now expected from the model: REVIEW, for genuinely ambiguous cases. The changelog for 1.0.14 is candid that this is a category-relative audit for the V5.3 workflow, opt-in, with older workflows unchanged.

Inputs and outputs

• image - source image. candidates - a DOGMA_CANDIDATES bundle.
• panel_size - 256–768, step 64, default 512.

Outputs are lists, index-aligned with each other:

• instance_sheets (IMAGE, list) - sheets of two candidates, same vertical two-up geometry as v1.0.11.
• audit_instructions (STRING, list) - one prompt per sheet, e.g. 1 PASS: short visual reason. numbering per tile.

Wire the sheet list and instruction list into a VLM together, then the returned text list into DOGMAReviewBatchV114. That reviewer wants exactly ceil(n/2) answers. It parses PASS, FAIL and REVIEW lines with a regex, keeps the short reason after the verdict (capped at 350 characters), strips a closed thinking block, and truncates the text hard if it finds an unclosed thinking tag. Contradictory verdicts for the same index become REVIEW rather than a coin flip, and any index with no readable line also becomes REVIEW - never a PASS.

Practical gotchas

Answer count is positional. The reviewer slices items in pairs, so if you accidentally buffer, filter or reorder the text list between the VLM and the reviewer, you get DOGMA: audit batch count mismatch. - which is the good outcome. The alternative would be audit verdicts silently applied to the wrong masks.

Reason text is the audit's real value. audit_report ends up lines like #3 solver/global/2/person/1: audit FAIL; REJECTED; mode=reject_explicit_fail; reason=selected sky above the roofline. Grep that string when coverage surprises you; don't just look at the retained mask count.

Zero candidates is not an error. You get a 1×64×64 sentinel image and No candidates., so an empty category slot can stay wired without exploding.

Install

Manager → search DOGMA Nodes, or:

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

requirements.txt ships scipy>=1.10; no weights are downloaded and no ComfyUI internals are patched. Around it, the pipeline needs SAM 3 (core comfy_extras SAM 3 detection - the KB flags Triton as a real Windows obstacle there) and a VLM node for the audit itself.

One structural note that applies to this whole node family: the versioned IDs are not vanity. DOGMAAuditBatchV114 internally calls the pack's v56.7 audit-view builder by looking its class up in ComfyUI's registry, and the v2 nodes explicitly raise DOGMA v2 requires the existing node pack containing <name> if a module fails to import. If someone trims the pack down to one file, newer nodes break in ways their own error message explains. Install the whole thing or none of it.

And keep the pack's honesty in mind: DOGMA Nodes comes from one author's DOGMA workflow, has no presence in the community discussion corpus, and its own notes say model calls in tests are mocked. v1.0.14's prompt logic is genuinely thought through; your VLM's judgement is the part you still have to check by eye.