Put a checkpoint between every stage.

A facade study reaches its final cleanup pass. The planting is convincing and the light is right, but three mullions moved during the second generation. Nobody notices until the image is placed beside the elevation. The last pass cannot restore geometry it never knew was approved.

Today's community sweep surfaced a three-stage architectural ComfyUI process: initial generation with ControlNet conditioning, additions through regional editing, then refinement. The useful idea is not the number three. It is the boundary between stages. Treat each boundary as a design checkpoint and later edits lose the right to rewrite earlier decisions.

Define the stages as contracts

A stage should name what it may change and what it must preserve. “First pass, second pass, final pass” describes sequence but not authority. “Structure, atmosphere, repair” tells the operator which evidence belongs in each branch and why an output can be rejected.

The checkpoint is a saved artifact, not a feeling. Export the accepted image, source input, workflow JSON, structural controls, seed, model identifier, and a one-line reason for approval. If a team member cannot reopen the approved state, there is no checkpoint.

Set the review conditions before generating. Decide the comparison resolution, monitor or print proof, named reviewer, and maximum turnaround. Keep the source visible during approval. A candidate viewed alone benefits from novelty; a candidate placed beside the model has to answer for every opening and edge. The review conditions should stay fixed through all three stages so the standard does not soften as the deadline approaches.

Stage one: establish structure

Begin with the most authoritative view available. A clean export from Revit, SketchUp, Rhino, or another model usually carries more dependable silhouette and opening information than a compressed presentation image. If the renderer can export depth, normals, object IDs, or linework, save those passes beside the source.

The official ComfyUI preprocessor documentation separates depth, lineart, pose, and normal extraction from generation. It says separate preprocessing supports faster iteration, clearer debugging, and reuse across passes. That separation fits studio review: inspect the control image before trusting the image it conditions.

Depth communicates relative distance. The official Depth ControlNet guide identifies architectural structure and perspective as a use case. Yet estimated depth is still an interpretation of pixels. Thin rails, glazing, deep shadows, and reflected facades may be misread. Compare the map with the source. Use linework where the boundary matters, depth where spatial order matters, and source passes where they are available.

The first acceptance sheet

Place source and candidate at identical size. Lower candidate opacity or use a difference blend. Check the roof profile, corners, slab edges, columns, window count, major joints, stairs, doors, and ground line. Mark each item pass or fail. A beautiful candidate with one extra bay remains a failed structure pass.

Do not introduce detailed planting or crowds yet. They hide the edges that need inspection and consume attention that belongs to the building. A plain first stage is easier to approve than a theatrical one.

When no candidate passes, change one variable at a time. Test the control signal alone, then prompt language, then generation strength. Keep failed outputs with their rejection notes. They reveal whether the graph repeatedly loses thin elements, confuses glazing with void, or bends a particular edge. A failure pattern is more useful than a lucky seed because it tells the team where manual protection is required.

Stage two: add atmosphere by region

Stage two starts from the approved structural file, not from an alternate first-stage candidate. Name the requested change: overcast light, late-afternoon warmth, wet pavement, two small trees, or occupied cafe seating. Put each addition inside an allowed region and list protected regions alongside it.

Planting needs clearance from entrances, accessible paths, signs, facade joints, and important sightlines. People need plausible scale and ground contact. Reflections should not erase mullions or invent interior floors. Weather may alter contrast and color, but it should not move the horizon or rebuild the roof.

Run additions in small groups. One pass can address light and sky. Another can add planting. A third can add people. This costs more queue actions, but it makes causality visible. When a doorway changes, the team can identify which branch did it instead of searching through one large prompt.

The second acceptance sheet

Reuse the structural checklist and add an intent check for each new element. Is that tree part of the proposed design or presentation dressing? Does the sky direction agree with cast shadows? Does a person block a required handrail? Can the same atmosphere be carried across companion views without contradicting the site?

Approve the atmosphere only after the structure still passes. Save a new file rather than overwriting stage one. The earlier approval is the recovery point when an attractive edit proves destructive.

For a set of views, appoint one frame as the atmosphere reference. Record sun direction, cloud cover, color temperature, planting species class, occupancy level, and wet or dry surfaces. The other cameras do not need identical pixels, but they should describe the same day and the same project. Independent prompting can produce five persuasive images that cannot coexist.

Stage three: repair named defects

The repair stage gets a defect list. Circle a broken chair leg, doubled person, noisy parapet, smeared plant edge, or banding in the sky. Each defect receives the smallest practical mask. The instruction should identify the correction and the surfaces that surround it.

Keep full-frame regeneration out of this stage. Upscaling can be useful, but a generative upscale can also redraw fine joints, lettering, fixtures, and faces. Compare the repaired output with the approved atmosphere file outside every mask. Any unexplained change is a new defect.

Review at delivery size as well as enlarged size. A hand that looks odd at 300 percent may be irrelevant in a small web image. A shifted vertical that disappears in a thumbnail may be obvious on a board. The delivery format decides which defect deserves another pass.

Keep a tiny stage register

The record can be a CSV, project note, or table in the image folder. Give every attempt an ID. Store parent ID, stage, workflow version, model, seed, controls, change request, rejection reason, reviewer, and status. This is not heavy project management. Ten fields can prevent a team from promoting the wrong image.

Use filenames that expose lineage, such as courtyard-v07-s1-approved.png, courtyard-v12-s2-approved.png, and courtyard-v15-s3-delivery.png. “final-final-2.png” hides the parent and the approval state.

A stage ends when its promises are checked, not when the queue stops.

Our take: iteration needs brakes

ComfyUI is good at keeping options open. Architecture work also needs moments when options close. Once the building geometry passes, later branches should treat it as evidence. Once atmosphere passes, repair should be local. Creativity remains inside each contract, while approved decisions stop drifting.

Start your next graph with three output nodes named Structure Review, Atmosphere Review, and Delivery Review. Connect nothing past a node until its file earns approval.

Save the image. Sign the stage. Then move.

Editorial basis: the 14 September 2026 ArchiGen AI intel sweep and its community result describing a three-stage ComfyUI architecture process. Claims about separated preprocessing, reusable control outputs, depth, lineart, and normals were checked against current official ComfyUI documentation. This is a workflow guide, not a claim of hands-on testing or a universal node recipe. ArchiGen AI carries no sponsored placements.