(AV) Scene
├─ shot:       example_townhouses/front   ← projects auto-discovered
├─ variation:  0..N  (increment to batch; every value = a unique combo)
├─ seed:       shuffles the variation order
└─ target_model: same dropdown as the Matrix
outputs: system_prompt · prompt · filename_tag · debug

Data lives in ComfyUI/user/archviz_scenes/ (examples copied on first run): templates/ (typology rules — townhouse_row / villa / tower), pools/ (global entourage lines, @guard: auto-appended), projects/ (one ~20-line file per project). Body text supports {field:…} and {preset:category/name} — the latter reads the same preset library as the Matrix, with per-model overrides honored.

New project = duplicate a project file, edit the fields, pick pools. New typology = write one template. Reproduce any render from its filename tag: type the variation + seed back in.

Target models (v5)

Pick your model in the (AV) Matrix target_model dropdown — the same preset stack compiles into that model's native prompt dialect:

| target_model | Output style | Use when | |---|---|---| | nano_banana_edit (default) | v4 edit instructions, byte-identical | Editing an existing render with Nano Banana / Gemini Image | | nano_banana_generate | Narrative prose | Text-to-image with Nano Banana | | flux2_pro | Natural-prose brief, word-budgeted | Flux 2 Pro (local or API) | | flux2_dev_structured | Labeled Subject/Scene/Camera/… | Flux 2 Dev / Flex / Klein | | gpt_image_2 | Photographer's brief + "Must not drift" block | GPT Image 2 | | ideogram4_json | Structured JSON caption (experimental) | Ideogram 4 |

Dialects are defined in presets/model_profiles.json — edit it to tune or add models, no code changes needed. Individual presets can carry per-model override text (see CHANGELOG 5.0.0).

Why this exists: v4 presets were written in Nano Banana's edit-instruction dialect. Fed to Flux 2 verbatim, "preserve the input image" instructions produce poor results — Flux wants a scene description, not an edit request. The compiler re-packages the same curated content per model.

A ComfyUI custom node package for high-end architectural visualization prompt construction with Nano Banana Pro / Gemini 3 Pro Image.

Designed specifically for image-to-image workflows where preserving input geometry is critical — the typical "screenshot or basic render → high-end render" workflow used by architecture firms.

What it does

Curated preset library for prompt construction across 10 categories
One Matrix node consolidates all preset selection (no canvas clutter)
Positive-constraint phrasing throughout (more reliable than negative prompts)
Photographer/equipment vocabulary in style presets (Hasselblad, Portra 400, tilt-shift, etc.)
Gulf Arab ethnicity refinement — Saudi, Emirati, Qatari, Omani specifics with appropriate dress
Firm-signature material presets — travertine + white textured plaster + mashrabiya combinations
Hero car options for luxury residential renders
Surroundings preservation with optional refinement levels

Quick start

Installation

Option A: ComfyUI Manager (once published to registry)

Search for "ArchViz Preset System" in ComfyUI Manager.

Option B: Manual git clone

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

Option C: Manual download

Download the latest release zip from Releases, extract into ComfyUI/custom_nodes/.

First run

Restart ComfyUI. On first launch, the bundled presets/default_presets.json is automatically copied to ComfyUI/user/archviz_presets.json. Your edits there survive future package updates.

You should see this in the console:

[ArchViz] First run — copied default presets to /path/to/ComfyUI/user/archviz_presets.json
[ArchViz] v3 loaded — 3 nodes ((AV) Preset / Matrix / Assembler)

Try the example workflow

Open workflows/render_variations_v3.json in ComfyUI. Drop a basic render or screenshot into the input image node and hit Queue. You'll get 4 distinct high-end variations in parallel.

The three nodes

All grouped under the ArchViz category in the right-click menu.

| Node | Use | |------|-----| | (AV) Preset | Single-category dropdown — for advanced/manual wiring | | (AV) Matrix | All 10 categories in one node — recommended default | | (AV) Assembler | Manual 6-fragment concatenator — flexibility |

The 13 categories

| Category | Purpose | |----------|---------| | preservation | Anchors geometry — most important slot | | camera | Composition, viewpoint, framing | | lighting | Time of day, light quality | | atmosphere | Weather, haze, conditions | | style | Photographic equipment + aesthetic | | materials | Material palette including firm signatures | | people | Figure presence with ethnicity options | | scale_life | Activity level + occupancy | | population_density | How many people and where placed | | hero_style | The compositional anchor figure's style | | cars | Hero vehicles for luxury residential (brand-free) | | surroundings | Context preservation with refinement options | | enhancement | Upscale, refine, or fix specific elements |

Pick (none) on any axis you don't want active. The Matrix node skips empty fragments.

The three-axis people system (v4.0+)

Scene population is split across three independent axes that combine to produce 11 × 4 × 6 = 264 figure variations:

people (ethnicity, no count) — emirati, saudi, qatari, omani, gulf_mixed, levantine_arab, european, afro_diaspora, east_asian, south_asian, mixed_diverse

population_density (count + placement) — single_hero (one foreground figure), hero_with_supporting (foreground hero + few background), multiple_active (several figures, no central hero), crowded_public (many figures)

hero_style (what the hero is wearing/doing) — elegant_woman_flowing_dress, elegant_man_tailored, contemplative_figure_from_behind, fashion_editorial_woman, couple_arrival_lifestyle, active_resident_lifestyle

The hero archetypes are based on the photographic tradition of Julius Schulman's iconic luxury residential photography (Stahl House, etc.) and modern high-end residential render conventions. The elegant_woman_flowing_dress and elegant_man_tailored archetypes work with any ethnicity — Western dress for Western ethnicities, abaya/thobe for Gulf ethnicities.

Recommended workflows for common use cases

Adding people to an existing finished render

When your input is already a high-quality render and you only want to add figures, use the minimal-change configuration to prevent softening of the rest of the image:

preservation → people_only_minimal_change
people → ethnicity of choice
population_density → single_hero, hero_with_supporting, or multiple_active
hero_style → archetype of choice (or (none) for non-hero figures)
All other categories → (none)

The people_only_minimal_change preservation preset uses compositing-style language that tells the model to add only the specified figures while leaving every other pixel of the input identical.

Generating variations of a render

When you want to explore different lighting, atmosphere, or style on the same building, use the standard preservation + descriptive categories:

preservation → preserve_design or preserve_with_polish
lighting, atmosphere, style, etc. → variations per branch

This is what the example workflow render_variations_v3.json demonstrates.

Aerial / masterplan shots

For drone-perspective and aerial views:

camera → aerial_oblique_45 (45° drone hero), aerial_high_overview (masterplan), or aerial_low_drone (cinematic)
atmosphere → aerial_perspective_strong (depth and scale) or clear_aerial_sharp (technical clarity)
surroundings → urban_dense_aerial_context, coastal_aerial_context, or desert_aerial_context to match the project setting
lighting → golden_hour or morning_soft are especially flattering from the air
population_density → typically (none) — figures are too small to read meaningfully at aerial scale

Transforming a screenshot or draft into a polished render

When the input is a SketchUp/Revit screenshot or rough render needing significant uplift:

preservation → preserve_design
style → photoreal_editorial or similar finished-look preset
materials, lighting, atmosphere → as desired

The model can do substantial transformation here because the input doesn't have fidelity to lose.

Enhancement workflows

The enhancement category covers a different mode of operation than the descriptive categories. Use it when you have an existing render that needs refinement rather than transformation:

Upscaling 1K → 4K: Use upscale_4k_full and set most other categories to (none). Set the Nano Banana resolution to 4K.
Vegetation refinement: Use enhance_vegetation_strict for fidelity, or enhance_vegetation_creative if you want richer planting.
Enhancing people holistically: Use enhance_people_full — refines anatomy, skin micro-texture, hair detail, clothing fabric, and posture together. This is what makes people look genuinely natural rather than just "fixed."
Polishing already-correct figures: Use enhance_people_polish — refines skin/hair/clothing surface quality without modifying anatomy or pose.
Surgical single-element fixes: Use fix_hands_only or fix_faces_only when only one part is broken.
Avoiding face/anatomy issues entirely: Use motion_blur_figures — a professional architectural photography technique that turns rough figures into elegant motion-blurred ghosts.

For enhancement runs, set most descriptive categories (lighting, atmosphere, style, etc.) to (none) so the prompt focuses on the refinement instruction. Stacking multiple enhancements in one pass can produce over-sharpened results — one clean pass beats three aggressive ones.

Editing the preset library

Edit ComfyUI/user/archviz_presets.json in any text editor.

{
  "category_name": {
    "preset_name": "the prompt text fragment"
  }
}

Rules:

Editing existing preset text: changes apply on next workflow run, no restart
Adding new preset names: restart ComfyUI to refresh dropdowns
Adding new categories: requires editing MATRIX_CATEGORIES in __init__.py

See CONTRIBUTING.md for guidance on writing good preset prompts.

Why image-to-image preservation matters

Nano Banana Pro / Gemini 3 Pro Image works best when the prompt describes only what's missing from the input — not the geometry, composition, and design that's already shown. This package is built around that principle:

The preservation category anchors the architecture so the model treats it as fixed
The system_prompt in the example workflow tells the model to render rather than redesign
Material/lighting/style fragments add the polish layer on top of the existing design

Compatibility

ComfyUI 0.20+ (subgraph support recommended for the example workflow)
WAS Node Suite (for the Image Save node in the example workflow)
ComfyUI's official Gemini Image node, OR any compatible Nano Banana / Gemini 3 Pro Image node

The custom nodes themselves have no dependencies outside the Python standard library.

Roadmap

LLM enhancer node integration (Gemini Flash) for user-intent rewriting
Per-axis blueprint subgraphs for advanced multi-pass workflows
Additional regional preset packs (Levant, Maghreb, South Asia specifics)

Contributing

See CONTRIBUTING.md. The most valuable contributions are well-written preset entries.

License

MIT — see LICENSE.