Choose a 1280-pixel clay or basic material render from a project you know. It should show a corner, repeated openings, glazing, a thin railing, and some background. Save the untouched source beside a folder called tests. Your goal is not a portfolio image. Your goal is to make one variable visible at a time.
Use a current checkpoint that your machine can run reliably, whether that is a Flux-family model or another image model supported by your installation. Pick one VAE, one sampler, one scheduler, and one fixed seed. Keep them fixed for the first three exercises. Model shopping can wait. If the graph cannot explain cause and effect, a newer model only produces a newer mystery.
Exercise one: prove the smallest graph
Build the shortest image-to-image chain your installation supports: load checkpoint, load image, encode image, positive and negative text conditioning, sampler, decode, preview, save. Avoid custom nodes. Avoid upscalers, detailers, LoRAs, adapters, and style packs. The first run should answer only one question: can this source image travel through the graph and return to disk?
Set denoise very low, around 0.10, and write a plain prompt describing what is already visible. Match the source dimensions if memory allows, or scale proportionally to a model-friendly size. Queue one image. Confirm that the camera, silhouette, window count, and crop remain nearly unchanged. Save both the image and the workflow embedded in it.
Now close ComfyUI, reopen it, drag the saved image onto the canvas, and run it again. A reproducible second run is the actual pass condition. If a missing node or model path stops it, fix that now. A graph that runs once in the browser session is a demonstration. A graph that reloads is a tool.
Exercise two: make a denoise contact sheet
Denoise strength is the most important early control because it governs how much authority the model receives to replace the source. Duplicate the same run at 0.15, 0.25, 0.35, and 0.50. Keep the seed, prompt, sampler, steps, and resolution identical. Put the four results beside the source.
Do not judge atmosphere first. Redline geometry. Count the window bays, inspect the roof edge, trace the corner, check railing spacing, and compare material boundaries. At low denoise, the result may look too close to the basic render. At high denoise, it may look convincing while quietly changing the project. Mark the first setting where a design fact moves. Your working ceiling is the step below it.
Repeat with one new seed only after finding that ceiling. If the threshold moves wildly between seeds, the workflow is not yet stable enough for production. Record the range instead of pretending one lucky image represents the setting.
A useful ComfyUI setting is not the one that made the best frame. It is the one whose failure boundary you can point to.
Exercise three: add one structural condition
Return to your chosen denoise ceiling. Add a single structural input, not three. If you still have the 3D scene, export a clean depth pass from the same camera. That is preferable because it comes from the actual geometry. If the model is unavailable, estimate depth from the source image and treat glass, thin rails, trees, and large flat walls with suspicion.
Load the depth image through the preprocessor or control path appropriate to your model. Begin with moderate control strength and apply it through most of the sampling range. Run the same fixed seed. Compare it with the uncontrolled version. The depth condition earns its place only if it protects spatial order without flattening materials or introducing halos.
Next, turn depth off and test an edge condition from the same view. Edge maps often defend window and roof boundaries more directly, but they can preserve unwanted linework and overstate every joint. Do not stack depth and edge yet. Place both results beside the uncontrolled image and name the defect each one fixes. Keep the condition that solves the defect your project actually has.
| Test | Change one thing | Pass condition |
|---|---|---|
| Small graph | Basic img2img chain | Reloads and reruns |
| Denoise sheet | Denoise only | Geometry ceiling is visible |
| Structure | Depth or edge only | Named defect improves |
| Local edit | Mask and prompt | Outside pixels stay still |
Exercise four: correct one bounded region
Architectural production is mostly correction. Make a mask for one simple area, such as the sky, a blank planting bed, or one wall finish. Feather the boundary slightly, but keep the mask clearly away from windows, roof edges, and railings on the first attempt. The prompt should describe only the replacement inside that region.
Run the edit at your known denoise ceiling. Then use a difference view in an image editor, or rapidly toggle between before and after. Look outside the mask. Any movement there is collateral change. Also inspect the mask boundary for repeated texture, smearing, or a visible seam. Reduce the affected region, lower denoise, or adjust mask behavior until the outside remains fixed.
Once the easy edit passes, move the mask against a real boundary: replace cladding next to a window, add planting behind a railing, or alter the sky around a thin parapet. This is where the graph becomes useful to an architect. It must understand that the requested change has an address.
Keep a one-page run log
Create a plain text or spreadsheet log with date, ComfyUI version, checkpoint filename, source dimensions, seed, sampler, steps, denoise, control type and strength, mask description, run time, and verdict. Add one sentence for the first visible failure. “Bay three became a door at 0.35” is more useful than “0.25 looked best.”
Save the workflow JSON with exact model filenames. Note every custom node repository only when you eventually add one. Test the final graph on a second view and, if this is a studio workflow, on a second workstation. A result that depends on an undocumented local install is not repeatable enough to hand to a colleague or return to after a deadline.
What not to learn yet
Do not begin with face detailers, tiled 4K upscaling, five ControlNets, style training, animation, or an elaborate manager full of custom nodes. Those can matter later. None teaches the first production skill, which is knowing why the building changed.
Do not chase the exact settings from a tutorial whose source image, GPU, model, and intended edit differ from yours. Copy its graph shape if it is clear, then rebuild the test on your own view. ComfyUI is valuable because it exposes the chain. Treating a downloaded workflow as a sealed appliance throws that value away.
Our take
The best first week with ComfyUI should produce four boring artifacts: a graph that reloads, a denoise contact sheet, a depth-versus-edge comparison, and a local edit with no collateral movement. Together they prove that you can run, measure, constrain, and correct.
Only then add polish. Keep the source fixed. Change one variable. Circle the first broken window.
Written from the 26 August 2026 intel sweep, which surfaced repeated community requests for practical ComfyUI architecture rendering tutorials and geometry-preserving enhancement workflows. ArchiGen AI carries no sponsored placements.