ComfyUI Node

ArchViz Controller

Turn a wall of prompt soup into six dropdowns

By Timmy-ComfyUI·Created 22 days ago·Updated 21 days ago· 3
ArchViz Controller
    • use_manual_prompt
    • environment
    • interior_lighting
    • interior_composition
    • interior_lens
    • exterior_lighting
    • exterior_composition
    • exterior_lens
    • landscape_lighting
    • landscape_composition
    • landscape_lens
    • seed
    • aspect_ratio
    • megapixels
    ◄use_manual_promptfalse►
    ◄environmentInterior►
    ◄interior_lightingSoft Daylight►
    ◄interior_compositionWide►
    ◄interior_lens24mm►
    ◄exterior_lightingOvercast Daylight►
    ◄exterior_compositionWide►
    ◄exterior_lens24mm►
    ◄landscape_lightingSoft Daylight►
    ◄landscape_compositionWide►
    ◄landscape_lens24mm►
    ◄seed0►
    ◄aspect_ratio2:3 (Portrait Photo)►
    ◄megapixels1.0►

    "Wide 24mm interior, soft daylight, editorial composition, warm bounced fill" is a sentence a renderer's client would nod at, and exactly the kind of thing a modern model eats up - but you don't want to retype eleven variants of it for eleven camera angles. ArchViz Controller holds that vocabulary as dropdowns, and it's built for one shipped workflow: the pack's T2I_KREA2_ComfyUI_ArchViz_v1.0.json, on Krea 2 / Krea 2 Turbo.

    Be clear on what it is: it generates no image and touches no model. It's a front-end for preset prompt text that lives inside the workflow file, and everything it outputs is a small integer.

    Why you'd actually want this

    The lens/lighting/composition dial isn't decoration. Krea 2's text encoder is Qwen3-VL, a general-purpose vision-language model, and on LLM-encoded stacks the highest-leverage prompt content is a named physical camera plus a stated lighting condition - phrasing the encoder was trained on. "24mm" is a stronger token sequence than "wide angle".

    So the real selling point is consistency: same lighting, same lens, same composition across a set of shots, chosen in one place instead of eight prompt boxes you swear you kept in sync. That's the archviz job - a client wants interior, exterior and landscape of one project that look like one photoshoot.

    Know the limit too: if your task is clay render → photoreal (ControlNet depth or canny over a Blender output), this isn't in that lane. It's text-to-image with a house style, not a render-finisher.

    How it works

    Eleven of the fourteen inputs are enums, and each one is converted to a 1-based index at execution. Interior → 1, Exterior → 2, Landscape → 3; lighting → 1–4; composition → 1–3; lens → 1–3. Those integers drive DazzleNodes' DazzleSwitch selectors inside the workflow's ArchViz Presets subgraph, where the actual prompt strings live in text nodes. The convention is documented in the source: 0 = normal/manual, 1 = input_01, 2 = input_02.

    One detail worth knowing because it will bite you: the lens ordering changes per environment. Interior and exterior are 24/35/50mm, so 2 = 35mm. Landscape is 24/50/85mm, so 2 = 50mm. The integer is just a pointer into whichever switch it lands on.

    The inputs and outputs that matter

    You'll set three things, honestly. environment picks the branch. use_manual_prompt decides whether your typed prompt stands alone or gets concatenated with the presets - in the shipped workflow it drives a Text Input Switch between your raw text and the combined version. Then aspect_ratio + megapixels set the output size: aspect_ratio is one of eight strings like 2:3 (Portrait Photo), megapixels is 0.1–16 with a 0.1 step, and both feed a core ResolutionSelector node which computes width/height rounded to a multiple of 64 for the latent. Note aspect_ratio comes out typed as *, which is what lets it plug into that node's combo widget.

    seed carries the usual control_after_generate dropdown; set widget control mode to Before if you want the number in the box to be the one that ran. The remaining fields - interior_lighting, interior_composition, interior_lens, and the exterior and landscape equivalents - are the preset pickers, and you only touch the trio belonging to your chosen environment. All three trios get exported on every run; environment is what tells the subgraph which one to actually use.

    Fourteen outputs, matching the inputs one for one: use_manual_prompt (BOOLEAN), the ten preset selectors (INT), seed (INT), aspect_ratio, megapixels (FLOAT). Nothing goes to a sampler directly.

    Installing it

    Through Manager, search ArchViz Presets & Controller (that's the registry display name - the repo title won't match). Or:

    cd ComfyUI/custom_nodes
    git clone https://github.com/Timmy-ComfyUI/comfyui_archviz_controller.git
    

    Restart, then open T2I_KREA2_ComfyUI_ArchViz_v1.0.json from the repo root.

    There are no Python dependencies - pyproject.toml declares an empty list and the module is pure stdlib. What you do need is the workflow's other four node packs, installed separately: ComfyUI-Custom-Scripts, WAS Node Suite (Revised), DJZ-Nodes, and DazzleNodes Collection. Skip those and the workflow loads with red missing nodes and the controller looks broken when it isn't.

    Where people get burned

    • You dragged the node into your own workflow and the presets do nothing. Of course they don't - the text lives in the subgraph, not the Python. Port the ArchViz Presets subgraph too, or the integers point at empty switches.
    • Missing node warning. That's the four packs above, not this one.
    • Model not found. The workflow loads krea2_turbo_bf16.safetensors, qwen3vl_4b_bf16.safetensors (CLIP type krea2) and qwen_image_vae.safetensors. Point the loaders at whatever you actually have. The KSampler is set to 8 steps at CFG 1.0 with euler/simple, which is the Turbo recipe - if you swap in the undistilled Raw checkpoint, that's a 50-ish step, CFG ~3.5 model and this sampler will not flatter it.
    • Megapixels above 1 at 21:9 or upscaled ratios. The selector is happy; your VRAM may not be.
    • Prompt folder empty. The LoadTextDirectory node's path and pattern are blank-ish by default - point them at a real folder, or ignore it entirely and type into the User Prompt node after disconnecting its input.
    • No support thread to find. As of writing there's essentially zero community chatter for this pack or its author, so the README and the Python source are your only documentation. It's a readable 200 lines, so that's fine - just don't go hunting for a Discord.
    CategoryArchViz

    Inputs (14)

    NameTypeDefaultDescription
    use_manual_promptBOOLEANfalse—
    environmentCOMBOInterior3 options: Interior, Exterior, Landscape
    interior_lightingCOMBOSoft Daylight4 options: Soft Daylight, Direct Sunlight, Warm Evening, Dramatic Directional
    interior_compositionCOMBOWide3 options: Wide, Editorial, Detail
    interior_lensCOMBO24mm3 options: 24mm, 35mm, 50mm
    exterior_lightingCOMBOOvercast Daylight4 options: Overcast Daylight, Direct Sunlight, Golden Hour, Blue Hour
    exterior_compositionCOMBOWide3 options: Wide, Editorial, Detail
    exterior_lensCOMBO24mm3 options: 24mm, 35mm, 50mm
    landscape_lightingCOMBOSoft Daylight4 options: Soft Daylight, Direct Sunlight, Golden Hour, Mist & Atmosphere
    landscape_compositionCOMBOWide3 options: Wide, Layered, Detail
    landscape_lensCOMBO24mm3 options: 24mm, 50mm, 85mm
    seedINT00–18446744073709550000—
    aspect_ratioCOMBO2:3 (Portrait Photo)8 options: 1:1 (Square), 2:3 (Portrait Photo), 3:2 (Photo), 3:4 (Portrait Standard), 4:3 (Standard), 9:16 (Portrait Widescreen), +2
    megapixelsFLOAT1.00.1–16—

    Outputs (14)

    NameTypeDescription
    use_manual_promptBOOLEAN—
    environmentINT—
    interior_lightingINT—
    interior_compositionINT—
    interior_lensINT—
    exterior_lightingINT—
    exterior_compositionINT—
    exterior_lensINT—
    landscape_lightingINT—
    landscape_compositionINT—
    landscape_lensINT—
    seedINT—
    aspect_ratio*—
    megapixelsFLOAT—