DOGMA v35 Object Settings — Defect-Only Prompt

Of the five settings nodes in this pack, v35 asks the least of anything. DOGMAObjectSettingsV35 takes v31's compact whole-object geometry, sets stitch_strength to 1.0, and replaces the per-crop instruction with a demand for a single defect, at most fourteen words, or the word none.

The output set

Two inputs, category and restoration_brief. Eight outputs, and the six geometry ones are inherited verbatim from v31: group_radius, context_px, max_source_side, target_long_side, max_crops - five crops flat across the real categories, wider proximity gaps (vehicles 120, people 96, faces 64, hands 56) - and stitch_strength at 1.0 for everything, because v35's stitch is an exact-core, opaque passthrough of the model's crop. The __none__ row still collapses to (48, 96, 1536, 1536, 1, 0.0): one crop, no stitch, no destruction.

vlm_request is where the version lives:

Inspect only the existing vehicles. Output exactly: DEFECT: followed by at most 14 words naming the single clearest visible defect. If no clear defect, output DEFECT: none. Do not describe the scene. Do not add/remove objects or change text.

Fourteen words is aggressive, and it's meant to be. Ask a VLM "what's wrong here" and you get a paragraph that names eight things, half of which are the model's imagination. Ask for the single clearest defect and you get one repairable instruction. The composer downstream strips the DEFECT: prefix, keeps only the first clause, truncates at ~120 characters, and rejects the answer outright if it contains remove/delete/erase/add/create/invent/replace/rewrite/rename/respell - on any of those, it emits PRESERVE {CATEGORY}. instead and the pipeline keeps the original crop exactly.

summary is your spec sheet. For vehicles it reads: v35 vehicles: v31 compact whole-object crops; Base9B 20-step CFG4 denoise0.35; Qwen defect-only ≤14 words; groups≤5; exact-core stitch.

How it fits the wider pipeline

SAM masks come in, a category planner picks slots, this node turns a category label into geometry plus a question, the VLM answers in one line, a prompt composer builds a short FIX … instruction, Klein runs the crop at denoise ~0.35, and a stitch node pastes the result back using the crop's recorded coordinates. This node is the smallest link in that chain and the one that most determines whether the local pass is surgical.

The tradeoff to understand: v35 is deliberately under-informed. It doesn't get the long universal/context/text-lock paragraphs that v27's request carried. It relies on the downstream prompt builder to keep count, position, identity and text intact, and on the safety gate to reject destructive wording. If you're restoring signage-heavy frames - shopfronts, plates, posters - that's a real risk: v35's prompt has less explicit "don't touch lettering" language than v27's, and the protection comes from the safety filter and your masks rather than from the instruction.

Install

ComfyUI Manager → search DOGMA Nodes, or:

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

Zero dependency install; the pack uses only PyTorch and ComfyUI's own APIs. The README documents the WAN VACE prep and sampler nodes only, so this node has no in-app help text - the source is 10,000 lines with every version retained side by side, which is why searching a version number in the node menu is the fastest way around the list.

Gotchas

Your VLM has to obey the format. If it answers with prose instead of DEFECT: …, the composer keeps the first clause and truncates, which usually still works - but a chatty VLM that leads with "I can see a car that appears to be…" gives you a prompt about a car being a car. Keep the answer short at the source.

DEFECT: none is a valid, common result. That's the node doing its job: no clear defect means the crop is preserved exactly rather than run through the model. If most crops come back preserved and you expected repairs, your SAM masks are probably catching objects that aren't actually damaged.

Watch the CFG. The author's stated defaults - base 9B, ~20 steps, CFG 4, denoise 0.35 - are for the base checkpoint. On Klein distilled, CFG 1 is correct and 4 will fry the crop.