Two questions appeared side by side in this week’s community sweep. One architect wanted a ComfyUI workflow to improve a render while preserving the design. Another wanted to learn ComfyUI for architectural work and asked whether any tutorials existed. Both questions describe a real need. Neither gives a workflow builder enough information to solve it.
Architecture makes generic ComfyUI advice especially risky. A portrait workflow can change a shirt button and remain successful. An architecture workflow can change a window bay and become false. The difference between a helpful graph and a screenshot of somebody else’s nodes is a compact, reproducible brief. Package the following six inputs before asking a forum, colleague, consultant, or model for help.
1. The actual source image
Post the image you intend to process, or a crop that preserves its technical characteristics. State where it came from: a SketchUp viewport, Revit export, Enscape frame, V-Ray render, hand sketch, photograph, or generated concept. Include the pixel dimensions and file type. A workflow for a clean 2048-pixel render is not automatically suitable for a 700-pixel screen capture with labels and compression artifacts.
If confidentiality prevents sharing the project, build a small proxy scene in the same application with the same export settings. A box, curtain wall, stair, tree, and reflective surface are enough to expose most workflow problems. Do not replace the source with an unrelated public image. The answer may work perfectly on that image and teach you nothing about your pipeline.
2. Every supporting pass you can export
List the geometry information available alongside the beauty image. Depth, normal, object ID, material ID, line, ambient occlusion, and cryptomatte passes each support different controls. You do not need all of them. You do need to tell the person helping you what exists.
For a model viewport, depth and line passes usually provide the most useful start. Depth controls volume and occlusion. Lines protect edges and openings. Material or object IDs make local changes far safer because they provide clean masks. If no passes exist, say so. The workflow may use a depth or edge preprocessor, but everyone should understand that those controls were inferred from pixels rather than exported from the source model.
“Keep the building” is a wish. “Keep 14 window bays, the roofline, and both stairs” is a test.
3. A written fidelity limit
Define what the model may change and what it must preserve. “Maintain the architecture” is too vague because architecture includes everything from massing to chair selection. Write two lists.
The protected list might include facade openings, floor plates, roof profile, structural grid, stairs, ramps, doors, fixed materials, and visible context buildings. The editable list might include sky, planting within designated beds, people, loose furniture, surface weathering, reflection intensity, and light quality. Mention the project phase. A concept-stage study permits more invention than a planning submission or client-approved design-development view.
Add a rejection rule that can be checked without taste. Examples include: reject if any window count changes, if a path becomes obstructed, if a material crosses an ID boundary, or if the silhouette differs from the line pass. This tells the workflow builder how much denoise and structural control the graph can tolerate.
4. Hardware and software inventory
State the operating system, GPU model, available VRAM, system memory, ComfyUI version or update date, and whether installing custom nodes is acceptable. If you use a cloud service, name the service and storage limits. A graph designed around a large Flux model may fail immediately on an 8 GB card. A graph dependent on five custom-node packs may violate a firm’s workstation policy.
List models already installed, including the exact checkpoint and ControlNet filenames. Names such as “SDXL” or “Flux” describe families, not runnable dependencies. Include the Python or CUDA error only when relevant, but always paste error text as text rather than only attaching a screenshot. Search and copy both depend on the exact wording.
5. One precise target output
Ask for a job, not a mood. “Make it realistic” combines lighting, materials, detail, context, camera, and taste. Start with one target such as relight an exterior from noon to overcast, replace placeholder planting without touching paths, add believable glazing reflections, or transfer a reference material onto one masked facade.
Specify delivery resolution, aspect ratio, and whether the result must match a set of other views. If consistency matters, attach one approved neighboring plate and say which qualities should match: color temperature, sky, planting species, camera treatment, or finish character. A single-image workflow can optimize for novelty. A project workflow must optimize for repeatability.
6. A reproducible package
If you already tried a graph, export the workflow JSON or a PNG containing the embedded workflow. Include the prompt, negative prompt, seed, sampler, scheduler, step count, CFG, denoise, model names, ControlNet weights, and preprocessors. Note what you expected, what happened, and the first node where output became wrong.
Remove client names, private paths, API keys, and embedded credentials before posting. Check whether the shared PNG still contains metadata you do not intend to publish. If a model or custom node has licensing limits, link to its original source rather than uploading a copy.
A minimal request can follow this template:
Source: 2048 x 1365 PNG from Enscape Passes: depth, material ID, line Protect: massing, 12 window bays, roof, entry ramp May change: sky, planting beds, reflections, light Machine: Windows, RTX 4070 12 GB, current ComfyUI Target: overcast exterior, same geometry, 2K output Tried: workflow JSON attached, SDXL checkpoint listed Failure: mullions merge above denoise 0.28 Reject if: window count or silhouette changes
How to evaluate the answer
A good response should identify dependencies, explain the role of each important node, and offer a small test before a large graph. It should separate model choice from structural control. It should also name the compromise: lower denoise preserves geometry but limits transformation, stronger edge control may harden unwanted lines, and larger models cost memory and time.
Be wary of answers that provide only a workflow image, omit model versions, or promise universal settings. Seeds and prompts do not travel cleanly across checkpoints. Custom nodes change. Preprocessors interpret drawings differently. The useful artifact is not a frozen node arrangement. It is a graph whose assumptions are visible enough to replace.
Our take: ask for the smallest proof
The fastest route into ComfyUI is not collecting a giant architecture workflow. It is proving one controlled transformation on one representative image. Ask for the smallest graph that can load your source, apply one structural condition, sample at a restrained denoise value, and save the result with metadata. Once that works, request masking, reference conditioning, tiled upscaling, or batch handling as separate additions.
This approach also makes community help easier to repay. When the graph works, post the corrected JSON, versions, settings, and a redacted before-and-after comparison. Describe the failure you fixed. The next architect searching the same error gets evidence instead of another unexplained download.
Attach the source, name the constraint, and make the first answer executable.
Written from the 23 August 2026 intel sweep, which surfaced community questions asking for ComfyUI architecture tutorials and workflows that preserve design fidelity. ArchiGen AI carries no sponsored placements.