(AV) Matrix

If you install this pack and use exactly one node, it's the (AV) Matrix. It's the whole point of the ArchViz preset system: thirteen dropdowns - preservation, camera, lighting, atmosphere, style, materials, people, population_density, hero_style, scale_life, cars, surroundings, enhancement - on a single node, plus a target_model dropdown that compiles the result into the native prompt dialect of whatever image model you're running. You pick your categories, it hands back a ready-to-paste prompt string. No generation, no API key, no GPU work; it's prompt construction for the image-to-image workflow where a sketch or screenshot becomes a high-end architectural render.

How it works

Every dropdown resolves to a hand-written fragment in ComfyUI/user/archviz_presets.json (auto-copied on first run; your edits survive package updates). Fragments assemble in a fixed order - preservation first because it anchors the input geometry, then camera, lighting, atmosphere, style, materials, people, scale_life, cars, surroundings, and finally your override text. That ordering is deliberate: preservation is the most important slot because it tells the model the architecture is fixed, not a suggestion. The whole library is written in positive-constraint phrasing, which is the reliable move on modern image models - describe what's there, don't lean on negative prompts.

The genuinely clever bit is target_model. The same curated fragments compile into each model's dialect: nano_banana_edit (the default, byte-identical to the pack's v4 output - edit instructions like "preserve the input image"), nano_banana_generate (narrative prose; drops preservation and enhancement since there's nothing to edit), flux2_pro (a word-budgeted natural-prose brief that strips edit-instruction sentences), flux2_dev_structured (labeled Subject/Scene/Camera slots for Dev/Flex/Klein), gpt_image_2 (a photographer's brief with "must not drift" constraints and hype words stripped), and ideogram4_json (a structured JSON caption). Feed the same selections to a different model and you get a differently-voiced prompt - the content stays yours, the phrasing adapts.

The inputs that matter

• preservation - the anchor. preserve_design / preserve_with_polish for variation work; people_only_minimal_change when you only want to add figures to an already-finished render.
• lighting, atmosphere, style, materials - your variation levers for exploring looks on the same building.
• enhancement - a different mode of operation: refinement instructions like upscale_4k_full, fix_hands_only, or motion_blur_figures. For these, set the descriptive categories to (none); stacking enhancements in one pass over-sharpens. One clean pass beats three aggressive ones.
• target_model - where the prompt is going. Change this instead of rewriting your selections.
• override (optional) - free-form additions, appended last. separator (optional, default ". ").

Outputs: prompt (the compiled string - wire into your image node's prompt slot) and filename_tag (a compact underscore slug built from your enhancement/lighting/style/people/cars picks, e.g. goldenhour_photoreal_editorial, for naming saved files).

Installing and troubleshooting

Same pack as the other ArchViz nodes - install once, get all four under the "ArchViz" category:

cd ComfyUI/custom_nodes/
git clone https://github.com/ArrowsVisuals/archviz-preset-system.git

Restart ComfyUI. No dependencies beyond Python's stdlib, no model downloads. ComfyUI Manager users search "ArchViz Preset System".

Two gotchas worth knowing. First, the pack doesn't run any model - Nano Banana Pro / Gemini Image, Flux 2, GPT Image 2 and Ideogram 4 each need their own node (the official Gemini Image node or a compatible Nano Banana node), and the closed ones need an API key and budget on top. The Matrix just writes them a better prompt. Second, don't bolt a big negative prompt onto the Matrix output and expect it to override the fragments - this pack is built on positive phrasing, and it's fighting itself if you are.