Map the defect before opening ComfyUI

The base render has five visible problems. The north wall is flat, the glass reads like black plastic, three shrubs are identical, the paving has no scale, and the crop makes the entrance hard to find. A common ComfyUI response is to connect a large enhancement graph and ask one prompt to fix everything.

That graph has no clean way to succeed. Lighting asks for a global tonal change. Glazing needs a bounded material correction. Planting needs local variation. Paving needs texture at a believable scale. Composition may require a new camera or outpaint. When those operations share one denoise value and one generation, the model trades one defect for another.

Today's community results repeated a familiar request: improve lighting, textures, reflections, vegetation, shadows, and composition while maintaining architecture. Turn that sentence into a defect map before loading a checkpoint.

Build the map from visible evidence

Duplicate the source image and mark every issue with a numbered circle. Use one noun and one failure: “glass, black,” “planting, repeated,” “paving, oversized,” “entry, unclear.” Avoid solution language at first. “Add cinematic sunset” is a proposed treatment, not a diagnosis.

For each mark, decide whether the defect is global, regional, structural, or compositional. Global defects affect the whole frame, such as weak contrast or a color cast. Regional defects sit inside a mask. Structural defects threaten geometry and require depth or edge guidance. Compositional defects often belong back in the 3D application.

Order the passes from broad to narrow. Correct composition in the model. Then establish global light. Repair materials and contextual regions. Finish with upscale and tiny cleanup. Every approved pass becomes the source for the next, while the untouched render remains available for comparison.

Fix composition outside ComfyUI first

If the entrance sits behind a column because of the camera, move the camera. If a tree blocks the facade in the model, adjust it there. Generative outpainting is useful when the designed view is sound but the delivery aspect ratio needs extra sky or foreground. It is a poor substitute for a deliberate architectural camera.

When outpainting is justified, extend one boundary at a time. Supply the source at full working resolution, mask only the new canvas, and protect the original image. Check perspective lines and repeated context where the seam meets the source. Never let an outpaint quietly become an alternate building elevation.

Give global light a small permission budget

For flat light, begin with img2img at low denoise, often around 0.15 to 0.30 depending on model, sampler, and source. Attach depth from the 3D model when possible. Estimated depth is a fallback, and reflective glass or thin rails can confuse it.

Use a short prompt that names light direction, softness, sky condition, and exposure. Hold materials and camera constant. Make a four-image contact sheet across a narrow denoise range with the same seed when supported. Choose the lowest setting that creates the required separation between facade planes.

Compare histograms only after visual inspection. A wider range is not automatically better. Check whether shaded interiors retain detail, white walls keep texture, and the sky remains subordinate. Most important, blink against the source and count the openings.

A pass is finished when its named defect is fixed and protected facts have not moved.

Treat glazing as architecture, not sparkle

Mask each glazing zone just inside the frame boundaries. Feather enough to prevent a hard pasted seam, but avoid mullions, slab edges, and door hardware. The prompt should describe transmission, reflection, interior visibility, and light condition. Do not ask for generic “realistic glass.”

Run exterior windows and entry doors separately when they have different depth or exposure. Composite the accepted pixels over the approved global pass. Inspect at 100 percent for doubled frames, invented rooms, reflected objects that do not exist, and inconsistent horizon lines. If geometry changes, shrink the mask, lower denoise, or add edge guidance.

Reflections must support the camera and surroundings. A dramatic mountain reflection in an urban courtyard is not added realism. It is new context.

Break planting into clusters

Repeated shrubs usually come from treating an entire planting band as one texture. Divide it into foreground, facade edge, and background clusters. Mask one cluster at a time and keep masks away from steps, doors, and low walls. Describe growth habit, approximate height, density, season, and regional plausibility.

Change the seed between clusters to reduce clones, but hold the general plant brief steady. Preserve some gaps. Real planting exposes soil, mulch, stems, and depth changes. A continuous green strip often hides the base of the building and makes an image look generated even when individual plants are convincing.

After compositing, run an edge inspection along the facade. Leaves should not cut through glazing or erase drainage details. Confirm that mature height does not contradict the planting plan. Generated vegetation is a presentation layer until the specified species and quantities are reconciled.

Repair material scale with a reference

For paving or cladding, crop a clean material reference that shows joints and surface character without unrelated objects. Mask the target plane, preserve its bounding edges, and use depth or line guidance when perspective is important. State an approximate module size in the prompt, but judge scale against known elements such as doors, risers, and facade bays.

Texture repair fails when it makes a flat sample look attractive while ignoring construction logic. Check joint direction at corners, course alignment, repetition, thickness, and transitions. If the model cannot hold them, return to the renderer and build the material procedurally. Generative correction should not replace information that the design team can model accurately.

Upscale only the approved composite

Do not upscale five candidates and hope detail resolves the choice. Flatten or version the accepted composite, attach the same depth and edge controls where the upscale includes a generative pass, and keep denoise low. Compare window counts, rail spacing, signage, faces, foliage edges, and fine joints with the pre-upscale image.

Save the workflow with checkpoint, VAE, sampler, scheduler, seed policy, input paths, mask files, control models, and custom-node versions. Name outputs by pass: 01-CAMERA, 02-LIGHT, 03-GLASS, 04-PLANTS, 05-MATERIAL, 06-UPSCALE. A teammate should be able to identify the last approved state without opening the graph.

The stop rule

After each pass, ask three questions. Did it fix the numbered defect? Did any protected fact move? Did it create a new defect larger than the original? Continue only when the answers are yes, no, and no. Otherwise revert, change one parameter, and test again.

Our take: the defect list is the workflow

A downloadable ComfyUI graph can show useful node connections, but it cannot diagnose the image in front of you. The durable method is to classify the failure, choose the smallest operation, define protected facts, and set an acceptance check before generation.

Start with a marked-up render, not a node screenshot. Make global changes once, local changes inside masks, structural changes under control guidance, and camera changes in the model. Keep every accepted state and discard attractive failures.

If the defect has no number, the model does not get to touch it.

Written from the 8 September 2026 intel sweep, which surfaced continued community demand for ComfyUI workflows that improve architectural renders while maintaining structure. Numerical settings are starting points and require testing with the selected model.