The render has doubled windows and flat materials. The beginner adds prompt words, raises ControlNet strength, installs another custom node, and changes the sampler. Four variables move. The next image is different, but nobody knows why.
Today’s sweep found the same community request in several forms: architects want a ComfyUI workflow that improves renders without losing composition. The useful answer is not another giant graph. It is a diagnostic order that tells you which kind of problem you have before you touch the controls.
Use four layers: installation, graph, conditioning, and architecture. A fault at an earlier layer invalidates conclusions about every layer after it.
Layer one: prove the installation
Start with the smallest official or known-good workflow for your chosen model family. Confirm that ComfyUI launches, sees the required checkpoint and VAE, accepts an input, completes one queue, and writes an output. Do not add architecture controls yet.
The official ComfyUI image-to-image guide illustrates this baseline. It names the checkpoint location, loads an input through the graph, and runs a KSampler with denoise below 1. Its purpose is modest: prove that image-to-image generation works. Treat that as a system test, not an architectural recipe.
If the queue stops, read the error before editing the prompt. Missing model files, unsupported node types, dependency conflicts, unavailable GPU memory, and incorrect paths are installation faults. A more detailed material description cannot repair them.
Installation pass condition
- The required models appear in their selectors.
- No node is marked missing or unresolved.
- The graph completes twice from the same saved setup.
- The output path and workflow file are known.
Save the working version, ComfyUI version, model filenames, and any custom-node revisions. That snapshot gives you somewhere to return when a later experiment breaks.
Layer two: prove the graph
A graph can execute and still be wired incorrectly. The wrong image may feed the latent encoder. A control model may load without reaching the positive and negative conditioning. An old branch may bypass the node you are adjusting. A resize node may change the aspect ratio before sampling.
Collapse the workflow until one path remains: load image, encode, prompt, sample, decode, save. Then add one control branch. Queue after every addition. If the output stops changing when a node is bypassed, that node may never have affected the active path.
Label inputs by job, not by file number. Use names such as source_render, depth_guide, edge_guide, and repair_mask. Match every control image to the final width and height. Preview intermediate images where the graph permits it. A depth map that looks inverted or an edge map filled with furniture clutter should be rejected before generation.
| Symptom | Likely layer | First check |
|---|---|---|
| Queue error or missing node | Installation | Models, dependencies, versions |
| Control slider changes nothing | Graph | Active links and conditioning path |
| Whole image drifts | Conditioning | Denoise and control signal |
| Convincing image, wrong building | Architecture | Openings, edges, levels, protected features |
Layer three: test the conditioning
Once the graph is known to work, freeze the installation and wiring. Use one source render and one seed where the workflow supports repeatability. Hold prompt, resolution, sampler, scheduler, model, and control image constant. Change denoise only.
ComfyUI’s documentation states the core relationship: lower denoise generally keeps the result closer to the input, while higher denoise permits more departure. That is a behavior to measure on your image, not a universal number to copy. Model families and source images respond differently.
Make a short denoise contact sheet across a restrained range appropriate to the task. Reject any setting that changes a protected opening, roof edge, stair, or camera. Select the lowest setting that produces the requested material or lighting change. Then freeze denoise and test control strength. Do not tune both in the same run.
If stronger control produces doubled edges, hard halos, or flattened texture, inspect the guide itself. Depth and line information may disagree. A railing that is visible in the edge map but absent from the depth map gives the model conflicting instructions. Turning both signals up makes the argument louder.
Change one cause, then name one effect.
Keep every candidate in a contact sheet with its seed and settings. Add a one-line rejection note: moved mullion, invented balcony, muddy glass, or no visible change. The rejected frames become a map of the workflow’s limits.
Layer four: inspect the building
A technically successful generation can still fail as architecture. It can render without errors, follow the prompt, and look polished while adding a floor, blocking an entrance, changing the window rhythm, or putting vegetation in an accessible route.
Write the architectural checks before generating. For an exterior, count levels and openings, trace the roof profile, compare primary edges, inspect entrances and rails, and verify the relation to visible context. For an interior, check doors, casework, fixtures, circulation width, ceiling features, and any approved material boundary.
Separate accuracy from appearance. First ask whether the image represents the design. Only then score light, material, atmosphere, and composition. If appearance is reviewed first, a beautiful failure gains emotional protection and survives longer than it should.
Use overlays or quick flicker comparisons against the source. They reveal shifted edges faster than side-by-side viewing. For a local correction, inspect outside the mask too. A masked workflow is only local if the saved output proves that unrelated regions stayed fixed.
A 20-minute diagnostic sequence
- Minutes 0 to 4: load the last known-good workflow and complete one baseline queue.
- Minutes 4 to 8: preview the source, latent path, and any control images. Confirm dimensions and active links.
- Minutes 8 to 14: run a small one-variable contact sheet, beginning with denoise.
- Minutes 14 to 18: compare the lowest useful setting against the architectural checklist.
- Minutes 18 to 20: save the accepted settings or record the failing layer and stop.
The stop matters. If the same protected feature fails twice after a controlled adjustment, do not keep feeding random seeds into the graph. Return to the source model, use a different control signal, isolate the region with a mask, or choose a conventional render path for that requirement.
What to ask when requesting help
A screenshot of the bad image and the phrase “my workflow does not work” forces everyone to guess. Provide the ComfyUI version, operating system, GPU and available memory, exact model filenames, workflow JSON, input dimensions, prompt, seed, sampler, scheduler, denoise, control weights, source image, control images, error text, and the architectural feature that failed.
State what changed immediately before the problem. If the graph has never completed, say so. If it completes but alters windows, say that instead. Installation support and design-fidelity support are different jobs.
This packet also protects the person asking. Advice can be tested against a fixed case, and a suggested change can be accepted or rejected for a visible reason. The conversation becomes diagnosis rather than folklore.
The opinion: complexity is not control
A graph with three preprocessors, two ControlNets, several LoRAs, an upscaler, and custom detail nodes may be powerful. It also creates more places for a failure to hide. For production work, add complexity only when a named defect survives the simpler path.
The community demand surfaced today is real. Architects do need better ComfyUI instruction. But the durable lesson is not which downloaded JSON happens to run this month. It is how to isolate a broken layer, return to a passing checkpoint, and make the next change legible.
Run the baseline. Trace the graph. Tune one condition. Count the windows.
Sources and method
This tutorial responds to repeated workflow questions in the 24 September 2026 community sweep. Technical behavior is grounded in the official ComfyUI image-to-image documentation, including its workflow-loading and denoise guidance. The diagnostic order is editorial guidance, not a report of a new hands-on benchmark. Exact settings must be validated on the reader’s installed version, models, and project files.