DOGMA Sector Settings v26.1 — Text Identity Lock

What it is

DOGMASectorSettingsV261 is DOGMASectorSettingsV26 with one paragraph appended to its vlm_request output. The numbers - group radius, context padding, source cap, target resolution, crop count, stitch strength - are byte-identical to v26. The per-category instructions are identical too. What's new is an explicit instruction block titled ABSOLUTE TEXT IDENTITY LOCK, and it's worth understanding because it tells you exactly which failure the author was losing shots to.

The problem it solves

Here's the sequence that ruins a frame. Your detailer finds the "architectural details" sector - windows, doors, balconies. In a real street scene, that sector's mask doesn't politely stop at the glass; it laps over the shop sign hanging in front of it. Your local VLM looks at the crop, writes a nice instruction like "clean up the storefront lettering, render the sign reading BAR CENTRALE in crisp white type," and hands that to FLUX.2 Klein. Klein, being a good instruction-follower, repaints the sign. Now your period-correct Italian advertisement has a sign that says something slightly different, in slightly different type, in slightly the wrong place. Nobody who knows the original can unsee it.

The lock forbids all of that. In the author's words: "all visible lettering, glyphs, numbers, logos, building names, rooftop letters, shop signs, posters, advertisements and license-plate characters inside or near this crop are fixed visual identity. Do not rewrite, rename, respell, replace, complete, infer or invent any text."

And then the line that does the real work: "If text is blurry or unreadable, preserve that same blurry/ambiguous glyph structure." Not "make it readable." Preserve the blur. That's the correct call for archival-style restoration, and it's the exact opposite of what a diffusion model wants to do with a smudged sign.

How it's wired

Same two inputs as v26: category (out of a sector plan node) and restoration_brief (multiline string) - both required, both forceInput. Note that in v26/v26.1 nothing from restoration_brief actually lands in the emitted request; the template is internal. It's there for graph threading.

Outputs are the same eight: group_radius, context_px, max_source_side, target_long_side, max_crops, stitch_strength, vlm_request, summary. The summary string gets a "| v26.1 TEXTSAFE: all text/signage locked" suffix, which is handy if you're running both versions side by side and want to know which one a text-display node is showing you.

The text lock works with the v26.1 planner, not instead of it. Remember what that planner does: any category whose name reads as signage is normalized to __none__ and never gets a generative sector at all, and "architectural details" is stripped down to windows, doors and balconies so storefront lettering can't sneak in with the mask. The settings-side lock is the second line of defence for text that ends up inside a crop anyway, because masks and boundaries in real photographs are never as clean as the intent.

Install

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

ComfyUI Manager → DOGMA Nodes works too. Restart. The pack's requirements.txt contains no dependencies at all - it's PyTorch and ComfyUI's built-in sampler APIs, no extra wheels, no model files, no keys.

Gotchas

• The Milan 1970s paragraph is still in there. v26.1 inherits v26's hard-coded historical context, appended to the lock. It's a production constant from the studio's own job, not a setting.
• The lock is a prompt, not a mechanism. It's a very strong prompt, but the model can still misbehave; verify signage on a real sample run before you batch a sequence.
• Don't pair this with a text-removal or text-upscaling node on the same crop and expect the lock to survive. The lock governs what the VLM writes, and if your instruction contains "render 'OPEN' clearly," you've undone it by hand.
• No community threads to compare notes with - the pack has effectively no reddit presence, and its README documents the WAN VACE prep nodes and samplers instead of this half. Read the docstrings; they're the manual.