A noon exterior arrives with flat concrete, empty planters, black glazing, and a washed sky. The brief says “enhance this render.” A single prompt tries to fix everything. The output returns at sunset, adds mature trees across the entrance, turns concrete into limestone, and invents a second floor behind the glass. It is more dramatic and less correct.
Recurring community posts in today's sweep ask for a ComfyUI workflow that improves architectural renders while maintaining the composition. The request often bundles lighting, texture, reflections, vegetation, and detail. Those targets should not enter one undifferentiated generation. Split them into jobs before choosing nodes.
Start with four routing questions
Write one sentence for each item: what may change, what must not change, what evidence exists, and how the result will be checked. “Make it photoreal” answers none of them. “Replace the gray facade finish inside the east elevation mask while keeping openings, joints, and camera fixed” answers all four.
Evidence can come from the source model, render passes, a flat image, a mask, or a reference. The source matters. A depth pass exported from the renderer represents scene geometry. Depth estimated from a JPG represents a model's reading of the pixels. The official ComfyUI preprocessor guide describes depth, lineart, pose, and normal signals as separate structural inputs that can be inspected and reused. Choose the signal whose information matches the edit.
| Job | Primary evidence | Protect | Reject when |
|---|---|---|---|
| Relight | Depth or normals, plus source | Geometry, materials, camera | Light creates false openings or moved shadows without intent |
| Material change | Object mask, edges, reference | Joints, openings, adjacent finishes | New material crosses the mask or changes form |
| Planting and entourage | Placement mask, depth order | Access, rails, sightlines, scale | People or plants block required elements |
| Cleanup and detail | Source crop and defect list | Everything outside named defects | Correction creates collateral change |
Job one: relight the scene
Relighting is about illumination and surface response, not redesign. Use the cleanest geometry evidence available. A source normal pass describes surface orientation and can help a relighting route understand which planes face the intended light. Depth establishes foreground and background order. ComfyUI's current documentation identifies normals as useful for relighting and structured edits, while its depth ControlNet guide positions depth as a way to control architectural spatial structure and perspective.
Keep appearance instructions narrow: time of day, direction and character of light, exposure intent, and sky condition. Do not ask for new materials or landscaping in the same pass. Compare roof edge, window count, column positions, and ground contact before judging mood. If the building changes, lower the permission to redraw, strengthen valid structural guidance, or return to a more authoritative pass.
Reflections need a separate note. A dark pane can be physically plausible without revealing an interior, and a bright reflection can conceal mullions in the source. Decide whether the task is to clarify the glazing, show the reflected context, or expose the interior. “Fix reflections” gives the graph permission to choose the answer.
Job two: replace a material
A material edit needs a boundary. Start with an object or material mask from the source renderer when possible. If only a flat image exists, draw the mask deliberately and inspect edges around reveals, frames, parapets, and vegetation. Feathering that looks harmless at screen scale can smear a finish across a window return.
Pair the mask with a reference that names the properties you want: module, joint width, roughness, color range, and weathering. “Warm stone” is a mood. “Light buff limestone panels, stacked coursing, narrow recessed joints, low gloss” is an instruction. Keep geometry controls focused on openings and silhouette rather than forcing every texture edge from the original image.
Run a difference check outside the mask. An attractive facade fails if the sky, neighboring wall, signage, or camera also changes. The current ChatGPT Images guidance notes that selected edits may extend beyond the highlighted area. That limitation is not unique evidence about ComfyUI, but it is a useful general warning: an edit boundary must be verified in the output, not assumed from the interface.
Job three: add planting and entourage
Planting is an occlusion problem before it is a style problem. Draw allowed placement zones and forbidden access zones. Use depth order to keep a tree behind a railing when it belongs behind it. Preserve door swings, accessible routes, facade signs, and sightlines. Give approximate height or human scale rather than asking for “lush landscaping.”
Separate permanent design elements from presentation atmosphere. Trees and hedges may represent a planting proposal. Pedestrians, bicycles, birds, and parked cars may only support scale and use. Label them accordingly in the review file. A generated oak placed over an entrance should not quietly become a design decision because it looks good.
Check repeated views when the image belongs to a set. A planting mass added independently to three cameras can occupy three incompatible locations. Either derive masks from the scene, use a shared spatial reference, or limit the output claim to one presentation frame.
Job four: repair, sharpen, and finish
Cleanup begins with a defect list, not a quality adjective. Circle a broken mullion, smeared chair, noisy shadow edge, malformed hand, or compression block. Route each crop or mask through the smallest correction that can solve it. Upscaling and restoration can sit after structural edits, but they should not become a final lottery that redraws the entire facade.
Inspect at both working size and delivery size. A defect visible at 200 percent may disappear in the client PDF. A window rhythm that looks acceptable in a thumbnail may fail on a presentation board. Save the pre-correction image, the mask, the output, and the setting change. If the repair cannot be reproduced, it is not yet part of a studio workflow.
One graph can contain four routes. One generation should not perform four contracts.
Build one router, not one giant prompt
A practical graph can share loading, color management, resolution, seed logging, and export nodes while branching into relight, material, planting, and cleanup groups. Keep the preprocessing stage separate, as ComfyUI's documentation recommends, so depth, lineart, and normals can be inspected and reused without rerunning everything.
Name every group by job and acceptance test. Bypass branches that are not needed. After each job, compare against the same protected-element checklist and approve a new base image only when it passes. The next route starts from that approved revision, not whichever output looked most cinematic in the queue.
Do not prescribe universal strength, denoise, sampler, or step values. Models and node implementations differ, and a number that protects one clay viewport can redraw a glossy interior. Record a tested range for the installed graph and show what fails at either end.
Our take: nouns beat adjectives
People asking for an architecture workflow do not need a larger tangle of nodes. They need the vague request divided into observable changes. Light has direction. A material has a boundary. A tree has a location and scale. A repair has a named defect. Each can be controlled and rejected on its own terms.
The word “enhance” encourages a model to decide what the image should become. Replace it with the job, the evidence, the protected elements, and the stop condition.
Do not enhance the render. Route the edit.
Editorial basis: the 13 September 2026 ArchiGen AI intel sweep and recurring community requests for ComfyUI architectural enhancement workflows. Technical roles for depth, lineart, normals, and separated preprocessing were checked against current official ComfyUI documentation. The selected-edit caution was checked against current official OpenAI help. This is a workflow design guide, not a claim of hands-on testing or a universal parameter recipe. ArchiGen AI carries no sponsored placements.