A facade has seven bays. The line drawing shows every mullion. The clay pass shows which planes step back, where the canopy casts shade, and how the stair meets the ground. Send only one of those files into an image workflow and the model must guess what the other one knew.
Today's intel sweep found the same practical advice recurring in architecture discussions around ComfyUI: begin with evidence from the 3D scene, including line drawings and simple-material views, then use conditioning to preserve structure while improving light, texture, planting, and atmosphere. Official ComfyUI documentation describes ControlNet as a way to guide generation with conditions such as edge and depth maps, and it documents chaining multiple ControlNet models. That makes a paired-input test possible. It does not guarantee that two controls will agree.
The useful question is not whether linework plus a clay pass produces a beautiful render. It is whether the pair protects more architectural facts than either input alone, at a cost the team can repeat.
Start with one locked camera
Save a named view in the modeling application. Record camera position, target, focal length or field of view, image dimensions, model revision, and visible option set. Export both passes at identical pixel dimensions without cropping or moving the camera.
The linework pass should be legible, not dense. Include the boundaries that matter to the test: silhouette, slab and roof edges, openings, major joints, stairs, and the ground line. Remove hatches, dimensions, hidden lines, grids, and minor seams that would turn the condition into visual noise.
The clay pass should explain volume. Use simple, distinct values for major materials or masses, a neutral background, and a fixed sun or ambient-light setup. Avoid reflections, detailed textures, people, planting, and decorative entourage. This is evidence about form, depth, and broad material regions, not a weak final render.
If the two exports do not align before generation, no node graph can make them reliable witnesses.
Overlay the exports at 50 percent opacity. Check corners, openings, horizon, and image bounds. A one-pixel antialiasing fringe is not the issue. A shifted camera, different crop, changed option, or altered aspect ratio is.
Keep the roles separate
Linework is good evidence for boundaries, repetition, and alignment. It is weak evidence for which plane sits in front, how deep a reveal is, or whether a dark region is material or shadow. A clay or simple-material view carries surface and depth cues, but soft lighting may hide a thin edge or repeated joint.
Do not assume that feeding both inputs automatically combines their strengths. The official ComfyUI guide notes that ControlNet models expect corresponding preprocessed reference images, that some preprocessors require custom nodes, and that control strength plus start and end percentages affect how a condition is applied. The chosen checkpoint and ControlNet must also be compatible. Record the exact models and preprocessing route rather than calling the graph “line plus clay.”
| Evidence | Ask it to protect | Watch for |
|---|---|---|
| Linework condition | Silhouette, openings, repeated bays, stairs, primary joints | False edges, doubled mullions, flattened depth, overdrawn texture |
| Clay or material image | Plane hierarchy, recesses, broad material zones, shadow logic | Lost thin elements, baked shadow treated as material, soft boundaries |
| Text prompt | Finish intent, time of day, atmosphere, named additions | Prompted objects replacing model facts |
| Mask, if used | Exact region permitted to change | Feather crossing protected edges, mismatched dimensions |
Run four cases, not one lucky queue
Build the smallest graph that can answer the question. Keep checkpoint, VAE, sampler, scheduler, steps, seed set, prompt, resolution, and source image constant. Change only the condition route across four cases.
- Clay only. Establish what the image route preserves without explicit line control.
- Linework only. Establish what edge guidance preserves and what spatial information it loses.
- Paired, equal starting influence. Apply both compatible conditions and record their strengths and active intervals.
- Paired, tuned once. Make one documented adjustment in response to a named failure, not a general search for the prettiest result.
Generate the same small seed set for every case. Four seeds per case is enough to expose whether a result was a repeatable tendency or one fortunate draw. Do not choose a different favorite seed for each route and call that a comparison.
Save the preprocessed condition images as outputs. A line extractor may not see the building the way the original vector export does. A depth or edge preprocessor may merge glazing, planting, shadow, and joints. The image that enters the control model is part of the evidence and belongs in the review.
Use a change map, not a beauty vote
Mark ten protected facts before running the graph. Pick facts that are visible and countable: seven facade bays, four stair risers, one continuous canopy edge, parapet height relative to the top opening, door position, column count, and the outline of the foreground slab. Add three permitted inventions, such as planting species, sky, and interior activity.
For every output, overlay the source linework and mark each protected fact pass, fail, or uncertain. Then record useful improvement separately: material readability, light direction, depth, glazing, vegetation, and removal of obvious synthetic artifacts. An output that wins the beauty vote but changes three protected facts has not passed an architectural enhancement test.
Also count the labor. Record preparation time for each pass, graph setup, generation count, review time, and any local repair. The paired route is worth keeping only when its extra preparation buys a measurable reduction in drift or correction work.
Tune conflicts in a fixed order
When the paired case performs worse, do not turn every control at once. First inspect alignment and preprocessing. Second remove irrelevant lines and muddy material separations at the source. Third lower one control's strength. Fourth shorten the interval during which that control acts. Only then consider a different compatible control model or checkpoint.
This order keeps diagnosis attached to evidence. If the canopy doubles because the edge map contains both a silhouette and a heavy shadow boundary, prompt changes will not solve the cause. If the clay pass already contains dramatic baked lighting, asking for a different time of day may create competing instructions.
Keep a rejected case when it explains a boundary. A failed paired run that shows line guidance flattening a deep recess can be more useful to the team than a polished output whose settings were never recorded.
Package the test for a second operator
Save the source view record, both original exports, every preprocessed control image, workflow JSON, dependency list, model identifiers, prompts, settings, seed list, outputs, marked change maps, and the decision. Ask another operator to rerun one selected case and identify the same protected facts.
The package should say what passed. “Use paired controls for exterior oblique views with repetitive openings” is a bounded result. “Use two ControlNets for architecture” is not. A later interior, night view, curved shell, or renovation overlay may need a different test.
Our take: export more truth, not more detail
The 3D model already knows more than a flattened beauty image can show. A line pass and a simple volume pass are two cheap ways to carry different parts of that knowledge into a ComfyUI experiment. Their value comes from agreement around one camera and one revision, not from the number of nodes connected afterward.
Official documentation supports conditional inputs, preprocessed references, adjustable control influence, and chained ControlNet models. It does not promise that a specific architectural pair will preserve a specific building. That proof belongs to the team, the saved view, and the change map.
Lock the camera. Export two witnesses. Make the pixels testify.
Editorial basis: the 19 September 2026 ArchiGen AI intel sweep, including current community discussion of ControlNet-based architectural visualization and advice to begin from 3D scene evidence; and the official ComfyUI ControlNet usage documentation covering conditional images, preprocessing, control strength, active intervals, and chained models. The paired-pass protocol, four-case test, protected-fact change map, and tuning order are editorial recommendations. Exact model compatibility must be checked for the installed workflow. This article does not claim hands-on testing.