The recurring ComfyUI question from architectural users is specific: how can a basic render gain convincing lighting, texture, reflections, vegetation, shadows, and composition without losing the building? The common answer is a large workflow graph. The more useful answer is an order of operations.
Architecture punishes uncontrolled improvement. A model that tries to solve every visual weakness at once has permission to reinterpret every pixel. The result may be richer, but windows move, edges soften, planting covers errors, and reflections describe spaces that do not exist. A controlled graph separates structural preservation from atmospheric and local finish work.
The four passes below are model-agnostic. Node names differ across installations, but the logic remains stable: prepare reliable controls, establish global light and material character, repair local semantic regions, then upscale under the same structural guidance.
Before generation: make three source maps
Begin with the cleanest viewport you can export. Do not ask diffusion to repair avoidable modeling or raster problems. Use a fixed camera and export at least a beauty or clay view, a depth map, and a line or edge image. A native depth pass is preferable because it knows the model. If only a flat image exists, a depth preprocessor can estimate one, but inspect thin canopies, rails, stairs, and layered glazing for errors.
Clean the edge image before connecting it. Canny often captures every material seam, tree twig, and compression artifact. Those marks become instructions. Reduce noise, remove lines that should not control geometry, and preserve the primary silhouette, openings, slab edges, mullion rhythm, and ground contact.
Create simple masks for glazing, vegetation, sky, and any facade material that needs special attention. The masks do not need to be beautiful. They need to prevent a local correction from spilling across an architectural boundary.
The control image is not a technical byproduct. It is the drawing you give the model permission to obey.
Pass 1: lock structure and composition
Use image-to-image generation at a conservative denoise setting. The exact value depends on model and sampler, but begin low enough that openings and silhouette remain stable. Attach depth guidance for spatial organization and cleaned edge guidance for boundaries. Keep the prompt literal: building type, principal materials, time of day, camera character, and the instruction to preserve geometry.
Do not request lush planting, cinematic fog, detailed interiors, wet paving, dramatic reflections, and tiny facade texture yet. Pass 1 has one task: replace the viewport's synthetic light and material response with a coherent photographic base while keeping the drawing legible.
Generate a small contact sheet using a fixed seed sequence. Reject frames with geometry errors before judging mood. Pick the frame with the strongest preservation among those that meet the basic light direction. A spectacular but altered design is not a base image. It is a detour.
Review gate after Pass 1
- Overlay the output on the source at partial opacity.
- Check silhouette, camera, opening count, slab edges, rails, and stairs.
- Confirm that material boundaries match the intended specification.
- Reject invented context that changes access, grade, or neighboring mass.
Pass 2: tune light and material response
Feed the approved base back into image-to-image at lower denoise. Keep depth and edge controls connected. This pass adjusts relationships across the whole frame: sky brightness, facade exposure, interior glow, shadow softness, glass reflectance, and the balance between foreground and building.
Prompt for physical relationships rather than adjectives. Specify overcast daylight with soft contact shadows, or late afternoon sun from camera left with interior exposure held below the facade. Ask for clear glass with restrained reflection and visible interior depth, not simply "realistic glass." Name the material finish, joint scale, and weathering level.
If the source render has a valid sun direction, preserve it. AI-generated highlights that contradict modeled shadows make an image feel wrong even when viewers cannot name the cause. Compare roof, mullion, canopy, and planting shadows as a set.
| Problem | Change here | Do not change |
|---|---|---|
| Flat facade | Light direction, material roughness | Window depth or bay count |
| Dead glazing | Reflection strength, interior exposure | Mullion spacing |
| Floating building | Contact shadows, ground tone | Finished floor level |
| Weak hierarchy | Exposure and local contrast | Camera or crop |
Pass 3: inpaint semantic regions
Now isolate the parts that need different instructions. Glazing, planting, sky, paving, signage, and interiors should not share one denoise decision. Use the prepared masks, feather them enough to blend, and process one region at a time. Keep the surrounding image visible as context.
Vegetation and ground
Planting must respect scale, climate, maintenance intent, paths, doors, and sightlines. Ask for the actual planting type and maturity rather than "lush landscaping." Keep denoise modest near building edges. Review trunks, roots, repeated plants, and shadows. Remove any branch that conveniently hides a failed corner.
Glazing and reflections
Mask glass inside the frame boundaries. Describe reflection strength, interior visibility, and the exterior context that should plausibly appear. Reflections should respond to view direction and sky, not insert a second unrelated building. Preserve mullions with edge guidance or exclude them from the mask.
Details and people
Add occupants and furniture only after the architecture is stable. Use scale references and separate masks for foreground and distant figures. Check hands less obsessively than circulation: no person should block a required route, stand inside furniture, or suggest access where none exists.
Pass 4: upscale without reopening the design
An upscale with denoise is another generative pass. Reattach depth and cleaned edge controls. Keep denoise lower than the stage that established the image. Tile only if memory requires it, and use enough overlap to prevent seams across straight lines and material fields.
Judge the upscale at actual delivery size. Excess microtexture makes concrete gritty, glass dirty, and foliage brittle. Compare it with the approved lower-resolution frame. The purpose is stable edges and credible fine detail, not a new interpretation.
After upscaling, perform ordinary image correction. Remove isolated artifacts, control color, apply restrained sharpening, and check values in grayscale. Save the pre-upscale frame, controls, seed, prompts, model versions, custom-node versions, and output settings with the project.
Build review gates into the graph
A graph can technically connect all four stages and run them automatically. Do not confuse automation with approval. Save an output after each pass and stop for review. The operator should be able to return to Pass 2 without regenerating Pass 1, or revise vegetation without disturbing the accepted facade.
Use bypass groups or separate workflow files if that makes review clearer. Label nodes by intent, not by package name: STRUCTURE, GLOBAL LIGHT, GLASS, PLANTING, UPSCALE. Another team member should understand what can change before touching a slider.
Maintain a rejection log for the first project. Record errors such as moved opening, doubled mullion, false reflection, broken stair, planting spill, tile seam, or changed camera. Patterns reveal whether a control weight, mask, preprocessor, or prompt is responsible.
Our take: smaller graphs can be stricter
The strongest architectural workflow is not the one with the most nodes. It is the one that makes responsibility visible. Depth governs space. Edges govern boundaries. Masks govern local permission. Denoise governs how much the model may reinterpret. Review gates govern whether the next step begins.
Start with one exterior and build only Pass 1. Prove that five seeds preserve the same opening pattern. Add Pass 2 and establish repeatable light. Then add one glass mask and one planting mask. Upscale last. Save the graph only after another person can reproduce the approved result from the same inputs.
One pass makes a picture. Four controlled passes make an image you can defend.
Written from the 1 September 2026 intel sweep, which surfaced repeated community requests for ComfyUI workflows that improve lighting, texture, reflections, vegetation, shadows, and composition while maintaining architecture. ArchiGen AI carries no sponsored placements.