A downloaded ComfyUI graph turns a bright SketchUp exterior into a credible brick building. The window rhythm holds, planting improves, and the late-afternoon light looks intentional. Then the team feeds it a dim interior. The depth preprocessor merges the glazing with the view beyond, lineart fills the ceiling with noise, and the saved denoise setting rebuilds the joinery.
This is the gap behind a recurring question in today's community sweep: where can architects find a ComfyUI workflow that enhances renders? A graph is easy to share. Its operating envelope is harder to see. The workflow may depend on a particular contrast range, camera type, image size, source cleanliness, model family, or undocumented custom node.
Before a studio calls any graph reusable, run a four-image portability test. The goal is not to prove that one setup handles every visual task. It is to identify the inputs it can accept, the conditions that require adjustment, and the point where a different graph should take over.
Define one job before testing four inputs
Portability only means something inside a fixed assignment. Write it in one sentence: “Improve material, light, and entourage in a model viewport while preserving camera, massing, openings, and floor count.” Do not let the task drift from enhancement to redesign as images change.
Set acceptance rules that can be checked across every source. Count visible openings. Mark the roof silhouette, floor lines, primary columns, stair edges, and horizon. Record the source dimensions and color space. Decide which changes are allowed. A stronger sky can pass; an added balcony cannot.
Lock the workflow version before the test. Save its JSON, checkpoint, VAE, ControlNet models, custom-node versions, prompts, seed policy, sampler, scheduler, step count, denoise value, and output dimensions. ComfyUI's Registry documentation says workflow JSON can identify registered custom nodes by globally unique names, which improves dependency tracing. It does not guarantee that an unregistered node, changed model, or missing asset will reproduce the result.
A portable graph is a documented range of acceptable inputs, not a screenshot of its best output.
Choose four images that expose different failures
Do not select four clean exterior views from the same model. That tests repetition, not portability. Use source images the studio actually receives, with one controlled challenge in each.
| Input | Challenge | Watch closely |
|---|---|---|
| Clean clay exterior | Simple geometry, weak material cues | Whether the graph invents facade detail |
| Glazed interior | Reflections, backlight, thin frames | Depth errors and disappearing mullions |
| Dusk viewport | Low contrast and bright apertures | False openings and lost silhouettes |
| Busy urban context | Trees, cars, signs, neighboring edges | Noise entering structural controls |
Keep the assignment identical. If the workflow is intended only for exteriors, replace the interior with a close facade crop or an aerial view. The important move is variation along relevant axes: contrast, edge density, depth ambiguity, scale, and source finish.
Inspect preprocessing before generation
ComfyUI's official preprocessor guide recommends separating depth, lineart, pose, and normals extraction into dedicated workflows. That separation makes outputs easier to inspect, refine, reuse, and debug. Use it. Export the preprocessor result for each of the four inputs before it reaches the sampler.
Compare depth maps at glass, overhangs, trees, and open sky. Look for collapsed planes, halos, and foreground objects that merge into the facade. Compare lineart at paving, shadows, vegetation, reflections, and repeated windows. A clean map on the clay exterior can become a black thicket on the urban view.
Do not compensate immediately. The first pass should use the saved graph unchanged. Portability is being measured, so every required edit counts. If one image needs a new threshold, manual mask, different preprocessor, or resized control map, log it as an intervention rather than quietly improving the setup.
Keep preprocessing artifacts as evidence
Name files by input, workflow revision, and stage. Preserve the untouched source, generated control maps, repaired maps, output grid, and settings record. ComfyUI notes that extracted depth and normal outputs can be reused across passes, which supports this modular record. It also lets a reviewer determine whether a failure began in source interpretation or generation.
Run the graph unchanged, then permit one repair
Generate a small fixed batch for each input with the same settings and a stable seed set. Four outputs per image are enough to reveal gross behavior without turning the exercise into a beauty contest. Score the required geometry before atmosphere or realism.
For any failed input, permit one named repair. It could be cleaning a control map, changing a preprocessor threshold, isolating a region with a mask, or reducing denoise. Run the same batch again. If the repair solves the defect without harming the other acceptance checks, it may become a documented branch in the workflow.
If the image requires several coupled changes, stop calling the original graph portable to that condition. Fork it. A glazed interior branch can use different preprocessing and generation settings while sharing file naming, review, and output conventions with the exterior branch.
Count intervention, not just successful images
A graph that eventually produces four good outputs may still be a poor studio standard if every new source needs expert diagnosis. Record setup minutes, manual edits, reruns, node substitutions, failed outputs, and reviewer time. Note which operator could complete each correction.
Classify each input as pass, conditional, or out. Pass means the locked graph met all acceptance rules. Conditional means one documented repair worked. Out means it changed protected geometry, needed an unbounded tuning session, or required a different technical route.
The resulting matrix is more useful than a universal “works for architecture” label. It might show that the graph passes clean daylight exteriors, is conditional for dusk, and is out for reflective interiors. The team now knows when to route an image elsewhere before spending generation credits.
Package the graph with an input contract
Put a short input contract beside the workflow JSON. State the intended job, supported source types, tested dimensions, required models and nodes, acceptance rules, known failure conditions, and approved repairs. Include one representative pass and one known failure.
Also state what the graph does not prove. A successful viewport enhancement does not validate a floor plan, measured daylight, construction detail, or material specification. The output remains a visual study that must be checked against the source model.
Retest after changing a checkpoint, ControlNet model, preprocessor, major custom node, or core ComfyUI version. Change one dependency at a time and rerun at least one easy input and one boundary input. The portability record belongs to a revision, not to the workflow name forever.
Our take: stop sharing naked JSON
The community is right to ask for practical architectural graphs. A visible node network can teach more than a prompt list. But the most valuable attachment is not another polished sample. It is the graph's tested input range and the record of where it breaks.
Swap the picture before you trust the pipeline.
Editorial basis: the 17 September 2026 ArchiGen AI intel sweep and recurring community requests for ComfyUI architectural enhancement workflows. Technical statements were checked against the current official ComfyUI preprocessor documentation and ComfyUI Registry overview. The four-image protocol is an editorial recommendation, not a claim of hands-on testing. Exact node behavior depends on installed models and extensions.