A colleague sends a gray SketchUp viewport and asks for a photoreal exterior. Another sends a hand sketch with one confident roofline. A third sends a nearly finished render that needs better planting and reflections. Searching for a single “ComfyUI architecture workflow” treats those three jobs as the same problem. They are not.
The graph should begin with an intake decision: what reliable information already exists in the source, and what is the model allowed to invent? A depth map protects volume. Edge guidance protects boundaries. A mask restricts the work area. A reference image steers appearance. Denoise controls how much of the source survives. Adding every control to every job creates conflict, extra tuning, and the false impression that a larger graph is a better one.
Today's community results again showed architects asking for a ComfyUI workflow that improves lighting, textures, reflections, vegetation, shadows, and composition while maintaining the building. That request contains several separate operations. Triage them before connecting a node.
Step one: name the source you actually have
Sort the input into one of four useful classes. A line sketch establishes composition and a few boundaries. A clay or shaded viewport establishes camera and modeled geometry. A basic render adds material regions and intentional light. A finished image already contains nearly everything and should receive local correction, not wholesale regeneration.
Do not promote the source in your head. A rough sketch does not contain precise window depth just because you know what the designer intended. A clay view does not define the final timber tone. A JPEG render may suggest depth, but its estimated depth map is weaker than a pass exported from the model. Write down what is known, what is approximate, and what may change.
| Source | Trust most | Primary control | Starting denoise |
|---|---|---|---|
| Line sketch | Composition and key edges | Cleaned Canny or line guidance | 0.55 to 0.75 |
| Clay viewport | Camera, massing, openings | Exported depth plus edges | 0.35 to 0.55 |
| Basic render | Geometry and material regions | Depth with selective masks | 0.20 to 0.40 |
| Finished image | Nearly all global information | Local masks and reference | 0.10 to 0.25 |
These denoise bands are starting points, not universal settings. Model family, sampler, resolution, and conditioning strength change the response. The principle is stable: sources carrying more approved information receive less permission to move.
For a sketch, protect the decisive lines
A hand sketch often contains doubled strokes, annotations, shadows, and construction marks. Feeding all of them into edge control can turn hesitation into architecture. Clean the input first. Retain the silhouette, floor lines, main openings, and the few edges that define the idea. Remove text, arrows, smudges, and incidental hatching.
Use one edge condition at moderate strength. Let the prompt and reference image establish materials and atmosphere because the sketch does not. Generate at a working resolution before upscaling. Review the silhouette and opening count, not the surface finish. If the massing moves, strengthen the useful edges or lower denoise. If every line becomes a dark physical seam, reduce edge strength rather than adding more prompt language.
For a clay viewport, export depth from the model
A modeled viewport knows more than a depth estimator can infer from its pixels. Export a depth pass from the same camera whenever the authoring or rendering application permits it. Normalize the useful range so the building is not compressed into a narrow band between sky and foreground.
Pair depth with a restrained edge pass when repetitive windows, rails, or roof breaks matter. Depth protects broad spatial order. Edges protect thin boundaries. They solve different failures. Begin with depth, add edges only when the first contact sheet shows details slipping.
Every control needs a named failure to prevent. If you cannot name the failure, remove the control.
Keep the camera fixed and ask for materials, weather, and occupancy in the prompt. If the model redraws openings, lower denoise before increasing every conditioning weight. High competing weights can produce brittle images with doubled edges or flat textures. Change one variable per contact sheet and label the output with its settings.
For a basic render, separate global and local work
A basic render may already have correct materials but weak light, empty planting, plastic glazing, and sterile context. Do not ask one img2img pass to solve all four. First decide whether the whole image needs a lighting pass. If it does, keep denoise low and preserve depth. Approve that result before touching local regions.
Then make separate masks for vegetation, glazing, foreground, or sky. Feather masks enough to avoid pasted edges, but keep them away from critical facade boundaries. Give each pass one sentence of instruction. “Add regionally plausible meadow planting with varied height” is testable. A long prompt containing planting, reflections, people, furniture, weather, and mood makes it impossible to know which instruction caused the wall to move.
Composite approved regions onto the approved global image. Do not continue regenerating the entire frame because one tree is wrong. The graph should narrow as the image improves.
For a finished image, choose a scalpel
If the source is client-ready except for one reflection, repeated person, or patch of vegetation, begin with a mask and the lowest denoise that produces a visible correction. Preserve an untouched copy beneath the generated layer. Compare the masked boundary at 100 percent and check whether color or grain shifts across it.
A reference image can help specify a plant type, material finish, or light quality. Crop the reference to the relevant evidence. A full mood image introduces composition and geometry signals you may not want. Reference strength should be high enough to carry appearance and low enough that the target remains the target.
Build an intake sheet before the JSON
For repeatable studio use, attach a one-page intake record to every workflow. List source class, approved facts, allowed changes, protected regions, target dimensions, model and VAE, control preprocessors, denoise range, seed policy, and acceptance checks. Link each input file by its project-relative path.
Save the graph only after a labeled low-resolution test passes. Embed or record the prompt and settings with the approved output. Pin custom-node versions if the workflow will leave your machine. A graph without its models, inputs, and version information is a screenshot of a process, not a transferable process.
A ten-minute diagnostic contact sheet
Generate four images using the same seed where the system supports it. Hold all settings fixed except denoise, tested at four points across the appropriate band. Put them beside the source. Choose the lowest value that makes the requested change. Then hold denoise and test control strength. This small grid tells you more than random prompts because each difference has one cause.
Circle failures directly on the sheet: moved mullion, merged chair, floating plant, erased rail, broken reflection. Those marks decide the next control. They also create a review artifact a colleague can understand without opening ComfyUI.
Our take: the smallest graph wins
Architecture tutorials often present a finished graph as the answer. The useful skill is not copying its nodes. It is reading the source and assigning only the controls needed to protect what the source knows.
Begin every job by classifying the input. Use edges for decisive lines, native depth for modeled space, masks for local change, and references for appearance. Set denoise according to how much approved information is already present. Add a node only after an output shows the failure that node is meant to correct.
Before downloading another workflow, look at the file on your desk. It already told you which graph to build.
Written from the 2 September 2026 intel sweep, which surfaced repeated community demand for ComfyUI workflows that enhance architectural renders while maintaining structure. Settings are starting ranges and should be validated against the chosen model.