Architectural visualization users keep asking how to use ComfyUI for final polish: better lighting, more convincing materials, improved reflections, credible planting, and stronger atmosphere without losing the building. Image-to-image processing can address those requests. It can also touch parts of the frame nobody asked it to change.
ComfyUI's official Stability AI image-to-image example describes an image denoise control from 0.0 to 1.0. Lower values keep the result closer to the input, while higher values move it toward text-to-image generation. That guidance explains the control, but a reviewer still needs to see what changed in a particular output.
Add a difference matte to every polished-render review. It is a visual map computed from the aligned base and candidate images. Unchanged areas fall dark. Changed pixels become bright or colored. The matte does not decide whether a change is good. It tells the reviewer where judgment is required.
The final image hides its own edit history
Put the base and polished render side by side and the eye goes first to the intended improvement. Warmer light feels better. Trees look fuller. Glass has more depth. That success makes small movement elsewhere harder to notice.
A difference matte reverses the priority. It makes every altered region compete for attention, including the quiet ones: a window edge displaced by three pixels, a handrail thinned against the background, a joint erased by new shadow, or a neighboring roofline absorbed into clouds.
The beauty image asks, “Do you like this?” The difference matte asks, “Did you authorize all of this?”
Build a four-view review packet
| View | Question | What it catches |
|---|---|---|
| Base | What facts entered the pass? | Source defects and approved geometry |
| Candidate | Did the requested improvement work? | Visual quality and communication |
| Difference matte | Where did pixels change? | Unrequested edits and broad drift |
| Blink or overlay | Did edges move? | Small alignment and silhouette errors |
All four views must share the same pixel dimensions, crop, and alignment. If the workflow resizes or shifts the candidate, register it to the base before calculating differences. Otherwise the matte will light up nearly every edge and conceal the changes that matter.
1. Freeze and preserve the accepted base
Use the base render that passed project review, not an earlier viewport capture or a compressed copy from a messaging app. Give it a stable filename and checksum. Record resolution, color space, model issue, camera, and approval date.
Do not replace this file when a new pass looks better. The accepted base is the factual reference for the session. If project geometry changes, create a new base revision and restart the comparison chain.
2. Define an allowed-change mask
Before generation, mark where change is permitted. A sky pass may include sky pixels and their reflection zones. A planting pass may include named planting beds and soft edge allowances. A material repair may include one facade surface while protecting openings, joints, signage, and adjacent materials.
The allowed mask is not the same as an inpainting mask, although it may begin from one. It is a review contract. Include transition zones where blending is expected, and exclude protected facts. Save it beside the base.
3. Generate the raw difference
Compute the absolute per-pixel difference between the aligned base and candidate. Display it with increased contrast so low-amplitude changes remain visible. Use a neutral background or a false-color palette that is distinct from the render.
Keep a raw numerical version as well as the presentation view. Heavy contrast makes review easier but can exaggerate harmless compression noise. The original difference values let the team distinguish widespread tiny variation from a smaller structural edit.
4. Separate allowed and unexpected change
Intersect the difference with the allowed-change mask. The part inside the mask is expected change and still needs visual review. The part outside is unexpected change and receives priority.
Use a threshold to ignore insignificant pixel noise, but set it from the delivery conditions rather than habit. A subtle shift may disappear on a small screen and remain obvious in a large print. Review several thresholds instead of letting one arbitrary value declare the image safe.
5. Blink the candidate against the base
A matte shows magnitude but not always direction. Alternate the registered base and candidate at a comfortable rate or drag a split overlay across them. Straight edges, repeated bays, columns, guardrails, curb lines, roof silhouettes, and horizon geometry reveal movement quickly.
Then inspect the candidate at delivery size and at 100 percent. A bright difference around every leaf may be completely appropriate. A faint continuous line along a protected facade edge may be a blocker.
Classify changes instead of chasing a zero matte
A useful polish pass is supposed to change pixels. The goal is not a black difference image. The goal is to make the distribution of change match the brief.
Classify each significant region as requested, necessary collateral, harmless incidental, or prohibited. Requested changes directly answer the brief. Necessary collateral includes small transitions required to integrate an edit. Harmless incidental change is visible but does not alter project facts or communication. Prohibited change touches protected geometry, graphics, specified content, or context.
Record the disposition for prohibited regions: lower denoise, add control, narrow or refine the mask, composite the base back in, use a local repair, or reject the candidate. Do not approve first and promise to remember the discrepancy later.
Use the matte to tune denoise with evidence
Denoise strength is often discussed as a creativity dial. For architectural polish, it is also a change-budget control. Run a small bracket at fixed seed and settings, then compare the candidate, difference matte, and outside-mask change for each value.
Select the lowest setting that achieves the named improvement while keeping prohibited changes below the review threshold. This will not produce one universal number. Different models, preprocessors, resolutions, prompts, and source images respond differently. The value becomes meaningful only beside the actual result and its change map.
If the desired improvement requires unacceptable global movement, stop raising the setting. Route the task to a tighter mask, a control input, conventional compositing, or a purpose-built renderer. A graph is not successful because it eventually produces a dramatic frame.
Package the evidence with the approved render
Save the base, candidate, allowed mask, raw difference, contrast-enhanced matte, workflow file, model and node versions, key parameters, seed, reviewer, and decision. Use filenames that link the set. A future editor should be able to reproduce the comparison without guessing which base was used.
For client delivery, the full packet may stay internal. For studio approval, it should be available at the decision point. The matte becomes especially valuable when several people review asynchronously, because it points each person toward the same changed regions.
Our take: every enhancement needs a visible boundary
ComfyUI gives architectural teams unusually explicit control over image processing, but a complex graph does not prove that the building survived. The proof lives in comparison.
Ask for richer reflections, calmer planting, or better light. Then put the candidate beside the accepted base, the allowed mask, and the difference matte. Approve the image only after the bright pixels tell the same story as the brief.
If the change has no map, it has no alibi.
Editorial basis: the 28 September 2026 ArchiGen AI intel sweep, including current community questions about using ComfyUI to improve lighting, materials, atmosphere, vegetation, and reflections, plus official ComfyUI documentation explaining image-to-image denoise behavior. This article reports no hands-on benchmark. The four-view packet, allowed-change mask, classifications, and approval method are editorial recommendations.