We pulled five older CGI renders from real client projects and ran them through an AI enhancement pass. The short version: AI made fabric, wood grain, and vegetation look noticeably more photographic. It also bent straight lines, garbled every piece of signage it touched, and quietly redesigned cabinet hardware nobody asked it to change. Not one of the five images was safe to hand to a client without manual correction.
Plenty has been written about AI and rendering, but most of it is theory. This post is a field report. Every observation below comes from an actual before and after pair, and you can inspect each one yourself.
Why We Ran AI Enhancement on Our Older Renders
Clients keep asking us a version of the same question: can AI refresh the renders we already paid for, so we do not have to pay for new ones? It is a fair question, and we wanted a tested answer instead of an opinion.
By “older renders” we mean images that are three to six years old. The geometry in them is still correct, because it came from CAD and Revit files. What ages is everything else: resolution, vegetation libraries, fabric detail, and post-processing style. A 2020 render often looks like a 2020 render, even when the design has not changed.
One thing this post is not: a general comparison of the two workflows. We already covered that ground in our CGI vs AI comparison guide. This is the follow-up experiment, run on our own archive, with results we can show.
How We Set Up the Test
We kept the AI render enhancement setup simple and repeatable, because a test you cannot repeat is just an anecdote.
The AI workflow we used
We ran each render through an img2img enhancement pass in InvokeAI, with ControlNet Canny edge guidance to hold the structure, followed by upscaling. We chose a locally controlled pipeline over one-click web upscalers for two reasons. Client files stay inside the studio, which matters for NDA work. And we can rerun the same image with the same settings, which matters for a fair test.
The five renders we picked
We deliberately picked five images that stress different things:
- A restaurant bar interior with mirrors, glassware, and heavy fabric
- A hospital corridor with a human figure, signage, and glossy floors
- A residential kitchen with wood, brushed metal, and stone in one frame
- A house exterior with detailed siding, a metal roof, and dense landscaping
- A family room with wood ceilings, a TV, and strong natural light
Between them, they cover most of what a working studio renders in a normal year.
How we judged the results
Our pass or fail standard was strict but practical: could a client approve this image without a correction round? We looked at texture realism, lighting and mood preservation, geometry accuracy, text and signage, and material or hardware fidelity. An image that looks better but shows the wrong cabinet pulls is not an improvement. It is a liability.
The Results: Five Before and After Comparisons
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Ut elit tellus, luctus nec ullamcorper mattis, pulvinar dapibus leo.
1. Hospital Corridor
Tested: AI staging (nurse insertion) + brightness/mood improvement.
What worked:
- Figure’s lighting matches the ambient illumination; pose natural
- Adds scale reference and life to a dead space
- Warmer balance kills the sterile CG feel
- Signage, door numbers, extinguisher, notice boards preserved — no text warping
What didn’t / needs manual fix:
- Floor reflection under the figure is weak for such glossy tile — reviewers disagreed on this, so treat reflection + shadow direction as a mandatory comp check
- Nurse scale plausible but verify against door heights
- Blue waste bin silently removed; slight halo around her silhouette at 100%
- Gloss level may read too polished for a real healthcare floor
Technical: strong use case for dropping figures into empty corridors, but AI edits the set without asking (removed bin). In healthcare visuals, signage, wall protection, and safety elements must be checked against the plan — small details matter here.
2. Florida Avenue Kitchen (Unit A12)
Tested: full restyle — material swap, relighting, restaging, detail enhancement.
What worked:
- Camera, layout, and cabinet reveals held with no warped geometry; new grey fronts rendered clean, no artifacts
- Lighting far more natural and less flat — bonsai now casts a believable wall shadow
- Backsplash, faucet, artwork, and side-room view survived the pass
- Decluttered island reads like real listing photography
What didn’t / needs manual fix:
- Warm wood cabinetry replaced with grey — complete loss of original design intent
- Plaster walls hallucinated into veined travertine; cooktop changed flush-electric → 5-burner gas
- Bar stool, place settings, and dishwasher panel deleted; track heads brass → silver; downlight count changed
- Microwave and fridge handles drift off-model
Technical: everything a client specs — materials, appliances, fixtures, furniture — drifted. If running a lighting/upscale pass, mask off cabinetry and appliances first. Mood exploration only; not safe for approval sets without CGI correction.
3. Stadler Exterior — View 05 Entry Court
Tested: rescue pass on a problem render — vignette, oversaturation, billboard plants, upscaling.
What worked:
- Heavy vignette and neon greens gone; balanced exposure with believable cloud sky
- Cobblestone path clarified — tactile, defined texture
- Foreground foliage gained crisp leaf detail; stronger light-through-trees mood
- Reads like a DSLR frame instead of a render
What didn’t / needs manual fix:
- Billboard-fern transparency ghosting still visible bottom-right — AI graded over the asset flaw instead of repairing it
- Shingle cladding flattened and roof edges went slightly wavy — classic hallucination, must be re-hardened manually
- Variegated foreground bushes morphed into generic foliage; intended species lost
- Small porch lamp added on the middle cabin; window reflections changed
Technical: pushed to enhance high-frequency detail (leaves, stone), AI sacrifices precise cladding and hard architectural edges. Asset-level problems (alpha-card ferns) need in-scene or paint fixes — grading won’t remove them.
4. Kennedy Center Art Studio
Tested: entourage generation, wall-art fill, upscaling, contrast grade.
What worked:
- Occupancy finally believable — varied, human-looking poses vs. stiff 3D figures
- Blank walls filled; strong working-studio atmosphere
- Grade fixed the washed-out haze; lighting on the easel field surprisingly consistent
- Skylights, clock, projector, yellow cabinet held position
What didn’t / needs manual fix:
- Original artwork replaced with invented paintings (one abstract turned into a red face) — content accuracy risk for a real facility
- Easel geometry merges/warps where frames overlap — some legs blend into each other and the floor
- Faces and hands go mushy at mid-distance; original figures replaced with new identities, not enhanced
- Floor changed light grey → dark polished — a material change, not a grade
Technical: biggest time-saver (entourage) and biggest failure (intersecting wooden geometry) in one image. AI can’t resolve dense overlapping frames — needs masking and CGI overpaint. Generated people and fictional art need rights/accuracy review before publishing.
5. Stadler Exterior — View 03 Skyline
Tested: relight, atmospheric clearing, vegetation and texture enhancement.
What worked:
- Biggest jump of the set — sky, canopy light, and ground read as photo
- Foreground bark gained convincing micro-detail; ferns and shrubs sharper and more natural
- Building more visible, image pulled out of the muddy “uncanny 3D” look
- Crude foreground ground plane rescued
What didn’t / needs manual fix:
- Facade drift: right gable turned brick/stone (was shingle + glass), mullion grids changed density, some glass reflections smudged
- Ground material changed gravel → asphalt uninstructed; road edge needs checking
- Background tree silhouettes went generic; planting layout no longer matches the landscape plan
- Massing reads slightly different — verify building count against the model
Technical: AI treats architecture as negotiable. Render a Material/Object ID pass to lock the building and let AI touch only forest and atmosphere. Landscape drift is tolerable; facade drift is not.
What AI Did Well in Every Test
AI earned its keep on organic material, atmosphere, and people. Those three strengths repeated across all five images.
- Organic textures. Fabric, wood grain, marble, bark, and grass all came back richer and less synthetic. The family room upholstery and the exterior planting were the clearest wins.
- Breaking CG repetition. Tiled textures are one of the classic tells of an older render. The AI pass added natural variation to wood floors and ceiling planks that would take hours to paint by hand.
- People. The nurse in the corridor scene went from an obvious 3D figure to someone who looks photographed in place. This was the single most dramatic improvement in the whole test.
- Mood preservation. We expected the passes to flatten our lighting. They mostly did not. The bar kept its warmth and the corridor kept its clinical white balance.
If your old render’s problem is that it feels sterile or dated, an AI pass addresses exactly that.
Where AI Failed Every Time
Every failure traced back to one root cause: the model does not know your building. Four failure types repeated across the tests.
- Geometry drift. Siding profiles, window mullions, cabinetry reveals, and wall seams all shifted. Small changes, but the kind an architect notices in seconds.
- Text and signage. The hospital signage became unreadable shapes that only resemble writing. This happened everywhere text appeared.
- Manufactured hard surfaces. Door hardware, hinges, chandelier structure, and a TV screen all warped or merged. AI treats a precise product like a texture to be reimagined.
- Physically correct light behavior. The bar mirrors reflected a room that does not exist, and sharp fixture falloff went soft and diffuse. The image looked plausible. It was also wrong.
Why this happens
An image model predicts plausible pixels based on patterns it learned from millions of pictures. It has no access to your floor plan, your product specs, or the physics of your lighting rig. CGI works the opposite way: the image is calculated from an actual 3D scene, so a mirror can only reflect what is really there. That is the whole difference in one sentence, and it explains every result in the table above.
What the Wider Industry Is Finding
Our results line up closely with what the broader visualization community reports, which is worth knowing before you build a workflow on either extreme.
In the CGarchitect AI survey of architectural visualization professionals, about 73 percent had already used AI tools in projects, and roughly 70 percent rated post-processing as AI’s most useful role. The same survey found the top complaints were inconsistent results, cited by about 77 percent, and lack of control over the final output, cited by 58 percent. Around 62 percent said AI was not ready for full production use. That is almost exactly our scorecard: strong finishing assistant, unreliable final authority.
Adoption keeps climbing anyway. The RIBA AI Report 2025 found that 59 percent of UK architecture practices now use AI, up from 41 percent the year before. The tools are becoming normal. The discipline around them has to catch up.
What Still Needs CGI
Anything that gets approved, purchased, or built still needs an accurate 3D scene behind it. Based on this test, we would not let an unmasked AI pass anywhere near the following:
- Approval images. If a client signs off against drawings or an FF&E schedule, the image must match them. Our architectural rendering services exist for exactly this standard.
- Product and furniture accuracy. A cabinet pull that AI rounded off is now the wrong SKU. For catalogs and ecommerce, that is a returns problem, not a style choice.
- Signage, branding, and any text. No exceptions. It failed in every instance.
- Consistent image sets. Ten views of one interior must agree with each other. AI passes drift image by image, which is why studios like ours keep interior rendering sets inside one controlled 3D scene.
- Animation, 360 tours, and VR. These need a coherent scene over time, not a single lucky frame.
- Revisions. “Change only the countertop” is a five minute edit in a 3D scene. In an AI pass, it is a new roll of the dice.
The Hybrid Workflow We Use Now
The test did not make us anti AI. It gave us a rulebook. Here is the production workflow we settled on after reviewing all five image pairs:
- Render the base image from the accurate 3D scene, as always.
- Render object ID masks for structure, hardware, signage, and screens.
- Run the AI pass only on organic zones: fabric, vegetation, skin, and atmosphere.
- Composite the protected CGI regions back over the AI output.
- QC the final image against the original geometry and the spec sheet before delivery.
What most teams miss is step two. Without masks, you are trusting a pattern generator with your client’s architecture. With masks, AI becomes what it is actually good at: a fast, tireless texture and atmosphere artist working under supervision.
Should You Enhance an Old Render or Re-render It?
Use this rule of thumb from our test results.
Enhancement with AI assist makes sense when:
- The design has not changed and the geometry is still correct
- The image is for marketing freshness, social content, or a resolution bump
- Nobody will approve, buy, or build from the image
- Text, hardware, and reflective surfaces can be masked or are absent
Re-rendering makes sense when:
- The design, materials, or SKUs have changed
- You need new camera angles, animation, or a consistent set
- The image faces approvals, buyers, or investors
- Accuracy is the whole point of the deliverable
Budget is usually the deciding factor, and the honest answer is that enhancement is cheaper per image while re-rendering is cheaper per mistake avoided. Our 3D rendering cost guide breaks down what drives price on the re-render side.
FAQ
Yes, within limits. In our five image test, AI reliably improved fabric, wood, stone, vegetation, people, and overall atmosphere. It also introduced geometry and text errors in every image, so the output needed manual correction before client use.
In our experience, always at least a little. Siding profiles, window mullions, cabinet hardware, and wall seams shifted in our tests even with edge guidance enabled. The changes are small but visible to anyone who knows the design.
We used InvokeAI with ControlNet for this test because it runs locally and is repeatable. Tools like Magnific, Krea, and Gemini based image editors do similar enhancement work. Every tool in this category shares the same core limitation: it predicts pixels without knowing your actual scene.
Only after review and correction. Treat the AI output as a draft, check it against the original geometry and spec, and replace anything structural that drifted. Unreviewed AI output is a risk to client trust, especially in architecture and product work.
Per image, usually yes, since the 3D scene does not need to be reopened. But correction time adds up, and a render that misrepresents the design can cost far more than it saved. Scope it case by case.
Be careful. Many web based tools process images on external servers, which can conflict with NDA obligations. This is one reason we ran this test on a locally hosted pipeline instead of a browser upscaler.
The Bottom Line
AI passed the texture test and failed the trust test. It is a strong finishing assistant for organic detail, people, and mood, and an unreliable narrator for geometry, text, and anything a client will measure. The practical move in 2026 is not choosing a side. It is masking the architecture, letting AI polish the environment, and keeping an accurate CGI scene as the source of truth.
If you have an older render library and are not sure which images can be refreshed and which need re-rendering, send a few of them over. We will look at the geometry, the intended use, and the risk points, then suggest the most practical path for each image. To see the quality bar we hold final images to, browse our 3D rendering portfolio.
