Two recurring requests appeared in today's community sweep. One user wanted a ComfyUI graph that could improve lighting, textures, reflections, vegetation, shadows, and overall composition. Another wanted an architectural render to become indistinguishable from a photograph. Both requests are understandable. Both combine several separate problems into the word photorealistic.
The same compression appears in comparison pages. Chaos evaluates “AI quality and control” as one of five criteria. Maverick Frame includes image quality among eight criteria while also warning that a polished frame can distort geometry and fail across a set. These guides correctly widen the decision beyond appearance. Yet a studio still needs a practical way to diagnose the image in front of it.
Photorealism is not a grade. It is a stack of claims that can fail independently.
Replace the verdict with a defect ledger
Start with the approved source view and the generated result at the same dimensions. Do not ask whether the result “looks real.” Record defects under six headings: geometry, materials, light, reflections, context, and image finish. Each entry should name a location, describe the mismatch, and state whether it blocks the intended use.
“Glass looks fake” is too vague. “North lobby pane reflects a bright sky that is absent from the visible exposure” is reviewable. “Plants are wrong” is not actionable. “Shrubs cover the accessible entrance and imply a planting bed outside the model boundary” gives the next operator a target.
| Ledger | Question | Typical evidence |
|---|---|---|
| Geometry | Did project form move? | Overlay, edge count, opening positions |
| Materials | Do scale and boundaries make sense? | Reference sample, joint spacing |
| Light | Can every major effect be explained? | Sun direction, fixture state, shadow |
| Reflections | Does reflected content belong? | View direction, adjacent surfaces |
| Context | Did the scene invent commitments? | Site plan, planting intent, access route |
| Finish | Does delivery expose artifacts? | Target crop and output size |
Geometry gets the first veto
A believable texture cannot rescue an incorrect opening. Before tuning style, compare fixed features that a reviewer can count: floors, structural bays, mullions, doors, stair landings, parapet breaks, and roof edges. Use a low-opacity overlay if the camera is intended to stay fixed. Mark every displaced or invented edge.
Classify each change as allowed, reviewable, or prohibited. A concept study may permit changes to cladding rhythm while protecting massing and floor count. A client image may protect every modeled opening. The category depends on the brief, not on how attractive the output is.
If geometry fails, return to the control stage. Adjust the input, mask, edge or depth guidance, or generation strength. Do not ask an upscaler to solve it. Sharpening the wrong window only makes the error easier to see.
Materials need scale, edges, and behavior
Material realism has three parts. Scale asks whether brick courses, panels, boards, and aggregate match the building. Edges ask whether one finish stops where the design says it stops. Behavior asks whether roughness, translucency, and reflectance respond plausibly to the scene.
Record those failures separately. A wall can have a plausible concrete surface but an invented control-joint grid. Timber can carry fine grain while every board runs across a corner without a joint. Glass can be glossy yet mirror content that cannot occupy the space behind the camera.
Use one source reference per critical finish, ideally a project sample or specified product image with known scale. The goal is not to copy a photograph. It is to give the reviewer a stable basis for saying why a generated surface passes or fails.
Light is a small physical system
List the visible sources before judging mood: sun, sky, interior fixtures, exterior fixtures, screens, and reflected light. Then check direction, softness, color, and occlusion. A warm room can still fail if a pendant does not illuminate the table below it. A dramatic facade can still fail if tree shadows point against the building shadow.
Separate exposure from illumination. Raising local brightness may reveal a facade, but it does not explain why that facade receives light. Adding contrast may make an image feel finished while hiding a source conflict. A good repair note names the source that should change.
For a controlled test, keep the camera and source image fixed. Change only one lighting instruction or one masked zone. Compare whether the requested region improved and whether unrelated regions moved. That second observation matters because a local request can produce a global rewrite.
Reflections are geometry in disguise
Review glazing, polished stone, water, metal, and wet pavement as views into other parts of the scene. Ask what each surface faces and what should be visible from that angle. Repeated trees, phantom buildings, doubled mullions, and incompatible skies are not minor texture defects. They change the implied site.
Mask reflection repairs narrowly and preserve the frame around them. Compare major reflected edges with nearby physical edges. If the tool repeatedly rewrites the glazing system, return to a conventional compositing or rendering step instead of spending more generations on a fragile patch.
Context can invent a design decision
Vegetation, people, furniture, vehicles, signs, and weather sell scale quickly. They can also cover entrances, erase accessible routes, imply nonexistent amenities, or alter a boundary condition. Treat them as project claims whenever the image leaves the internal concept review.
Create a context allowance before generation. It can permit generic distant planting while protecting the first five meters around the building. It can allow anonymous figures while prohibiting legible brands and safety equipment changes. It can allow atmospheric wet pavement while requiring that drains and thresholds remain readable.
Every context defect should identify whether to remove, replace, or relocate an element. “Less busy” is a taste note. “Remove the tree blocking the west entrance” is a repair instruction.
Judge finish at the delivery size
Image finish includes edge halos, repeated microtexture, waxy people, noisy foliage, oversharpening, banding, and compression. Inspect once at working size and once at the actual presentation crop. A defect that vanishes in a small internal pin-up may still matter in a full-page proposal. A pore-level oddity that disappears at delivery size may not deserve another pass.
Keep finish last because it is the easiest category to improve while earlier errors remain. Crispness creates confidence. It does not create accuracy.
Our take: count closed defects, not prettier versions
Run one baseline, create the ledger, and assign each defect a priority and an owner. For every repair pass, record the target, method, seed or settings, result, and collateral changes. Stop when all blocking defects are closed, not when the team runs out of adjectives.
This method does not promise that ComfyUI, a BIM plugin, or a conventional renderer will win. It reveals which stage owns each problem. Geometry may need a stricter source. Materials may need a local mask. Reflections may need compositing. Finish may need an upscale or manual retouch. The ledger routes work without pretending one control fixes everything.
Do not ask whether the render became more real. Ask which lie was removed.
Editorial basis: the 1 October 2026 ArchiGen AI intel sweep; two community requests about enhancing architectural renders in ComfyUI and making an architectural render resemble a photograph; and the published evaluation criteria in the Chaos six-tool guide and Maverick Frame comparison. Sources were checked on 1 October 2026. Community posts establish user questions, not validated workflows. Vendor and studio pages establish their stated methods and claims. This article reports no hands-on test. The ledger is an editorial review protocol.