Architectural render enhancement needs a punch list

This week’s community sweep surfaced a familiar request: take an architectural render into ComfyUI, improve lighting, textures, reflections, vegetation, shadows, and composition, while maintaining the design. The request sounds reasonable because every item is ordinary post-production work. Combined in one diffusion pass, though, those six instructions conflict. Composition asks for broad change. Design fidelity asks for almost none. The model cannot know which pixel belongs to the architect and which belongs to the presentation.

The useful workflow is therefore not one magic graph. It is a sequence of small passes governed by a punch list. Each pass gets one visual job, one protected region, and one reason to reject the result. That is slower than pressing Queue once and much faster than discovering, after a client comment, that the attractive second-floor shadow hides a balcony the model deleted.

Start with a fidelity contract

Before opening ComfyUI, write down what cannot change. Count facade bays. Mark roof edges, slab lines, window openings, guardrails, doors, signage, and any visible accessibility feature. If the image belongs to a real project, add material boundaries that correspond to the finish schedule. This becomes the fidelity contract.

Export three supporting passes from the source model when possible: depth, object or material ID, and a clean line pass. Depth holds the major volumes. IDs make accurate masks. Lines expose drift that a glossy output can hide. If the only input is a flat image, preprocess depth and edges, but treat them as estimates. A recovered depth map is useful guidance, not a substitute for the model that produced the view.

One prompt describes an ambition. A punch list describes work that can be checked.

The order matters more than the model

Run the passes from broad environmental conditions to small local corrections. Large lighting changes affect every later judgment. Fine texture and planting should come after the light is stable. Upscaling comes last because it multiplies whatever mistakes remain.

Pass one: establish light without changing mass

Lighting is the only pass that may reasonably touch the full frame. Keep denoise low and structural guidance strong. Use a depth or line ControlNet, then prompt for a specific condition: overcast north light, late-afternoon side light, or soft interior light at blue hour. Avoid style bundles such as cinematic, dramatic, award-winning, and photoreal. Those terms encourage the model to solve the whole image rather than the illumination.

Compare the output with the source as a difference blend in Photoshop or another editor. Check the silhouette first, then window counts, then horizontal slab edges. If those move, lower denoise or increase structural control. Do not accept geometry drift because the shadows look good. Good shadows can be generated again. An untracked design change can survive into every later plate.

Pass two: correct one material family at a time

“Improve textures” is too broad. Concrete, timber, stone, metal, and paving fail differently. Mask one family using the material ID pass, feather the mask slightly, and describe scale as well as finish. Board-formed concrete needs formwork spacing. Brick needs bond and believable unit size. Timber needs grain that follows the member, not a generic brown texture floating across it.

Keep original joints and boundaries visible through the process. Diffusion is good at plausible surface information and bad at remembering that a reveal aligns with a detail elsewhere in the set. If a finish is documented, use a reference image and restrained image conditioning. The goal is to restore material cues that the base render lost, not ask the model to select a nicer product.

Pass three: treat glazing as an optical system

Glass is where “more realistic” often produces the loudest fiction. A model may paste a skyline into a reflection, brighten an interior that should be dark, or turn a curtain wall into a mirror. Mask glazing while protecting mullions. Ask for reflection strength, interior visibility, and roughness appropriate to the scene. Those are separate variables.

Reflection direction should agree with what stands behind the camera and across the street. Shadow and sky conditions should agree with the main light pass. For a multi-view set, use the same environmental reference and prompt language across every view. If each plate receives a different invented city, the building looks as if it moved between images.

Pass four: add planting with circulation protected

Vegetation benefits from generative variation, but it must remain subordinate to the site plan. Use masks for planted zones, then create negative masks over paths, ramps, entrances, fire access, and sight lines. Describe plant habit and scale instead of requesting “lush landscaping.” A low ornamental grass, a clipped hedge, and a mature canopy tree have different effects on visibility and spatial reading.

Run foreground, middle distance, and background planting separately if the scene needs substantial work. This prevents a single pass from turning every green pixel into the same species at the same sharpness. It also makes rejection cheap. A bad foreground pass can be discarded without losing approved trees in the court.

Pass five: local cleanup, not a second redesign

Now inspect the image at 100 percent. Make a list of defects: repeating leaves, broken people, melted chair legs, false joints, smeared signage, doubled mullions. Inpaint each small region with surrounding context and modest denoise. Save after every accepted group of corrections. A giant cleanup mask reopens parts of the image that already passed review.

This is also the moment to decide whether AI is still the right tool. Straighten a line with a conventional transform. Repair a letter with type. Clone a small texture seam. Diffusion is useful when the missing content is ambiguous. It is unnecessary when the correct answer is already known to the pixel.

Composition is a design decision, so separate it

The original community request included composition. That item does not belong in an enhancement graph. Cropping, camera height, focal length, and view direction determine what the project communicates. If the composition is weak, return to the 3D camera and export another base view. Generative outpainting can test whether more sky or foreground might help, but it should not become undocumented site context.

For a speculative concept image, broader change may be acceptable. Label it as concept work and save it outside the approval sequence. For a project render, “improved composition” cannot quietly mean a taller lobby, wider court, or missing neighboring wall.

Our take: queue fewer changes

Architects searching for a ComfyUI enhancement workflow are often handed graphs with dozens of nodes and no review logic. Node count is not control. Control comes from dividing the image into decisions a person can approve. Build the first version with image load, structural conditioning, sampler, mask handling, and save nodes. Add complexity only when a named defect requires it.

Run the workflow on one difficult view before processing the full set. Record denoise, control weight, seed, model, and prompt for every accepted pass. Then repeat the approved recipe on an adjacent view and look for inconsistencies. The test is not whether ComfyUI can make one image attractive. It is whether the method can improve twelve plates without losing the project between them.

Open the source render, circle the first defect, and fix only that. The building stays yours.

Written from the 23 August 2026 intel sweep, which surfaced community requests for ComfyUI workflows that improve lighting, textures, reflections, vegetation, shadows, and composition while maintaining architectural design. ArchiGen AI carries no sponsored placements.