The timer on screen stops at 28 seconds. The first image has a luminous lobby, mature planting, and a roof that no longer meets the core. Three reruns follow. Someone masks the windows, repairs the parapet, exports a larger source, waits through another queue, and asks the project architect which version still represents the scheme. The image generator was fast. The job was not.
This distinction matters because speed is now a primary sales argument for AI rendering. Today's intel sweep surfaced a community claim that small firms can upload a SketchUp screenshot and get a solid result in under 30 seconds, alongside several 2026 tool comparisons built around faster visualization. The claim may describe a real generation interval. It does not describe production time.
A studio evaluating a tool needs a clock it can defend. The test below measures one route from a prepared design view to an image accepted for its stated use. It does not assume that an AI image is a construction document, and it does not pretend that every output needs final-render precision. It asks one simpler question: how long did this deliverable actually occupy the team?
Define the finish line before the start
A stopwatch is meaningless if the tester can stop when an image feels impressive. Write the intended use first. A concept pin-up may accept changed planting and approximate material joints. A planning submission cannot casually alter floor count, massing, access, or neighboring context. A client-facing design-development image may require exact openings, specified finishes, and consistency with other views.
Turn that use into five acceptance checks. Keep them visible beside the tool. A practical set is: major geometry preserved, opening count and position preserved, camera unchanged, required materials readable, and prohibited inventions absent. Add project-specific checks such as accessible route, signage, guardrails, or retained trees. The timer stops only when an output passes every required check.
This follows the boundary described in Chaos's current guide to what AI rendering can and cannot automate: generative speed is most useful in concept work, while precision and professional responsibility become stricter as a project advances. One speed number cannot cover both jobs.
Use six clocks, not one
| Clock | Start and stop | Why it counts |
|---|---|---|
| Preparation | From opening the source view to a valid upload or live viewport | Hiding guides, fixing exposure, choosing resolution, and exporting are labor. |
| Instruction | From first brief entry through reference and control setup | A fast model with a fragile prompt can move effort upstream. |
| Queue and generation | From submit to visible result | This is the interval most demos report, and it is still worth recording. |
| Selection | Time spent comparing, rejecting, and discussing candidates | A batch of four creates four inspection tasks. |
| Correction | All reruns, masks, local edits, compositing, and retouching | Correction is often where a quick first image becomes expensive. |
| Approval | Final architecture check and export for the stated use | An attractive image is unfinished until an accountable reviewer accepts it. |
Record elapsed time and active labor separately. A five-minute cloud queue may block no one. A five-minute correction may require continuous attention from a senior designer. Both affect deadlines, but only the second belongs fully in labor cost. Also record the role performing each active task. Ten principal minutes do not cost the same as ten unattended compute minutes.
The honest number is not time to first image. It is time to first acceptable image, for a use named before the test.
Run an ordinary view, not a hero shot
Choose a view that represents weekly work and contains enough difficulty to expose failure. Include repeated windows, a roof edge against the sky, glazing, a foreground object, planting, and one material junction. Avoid the studio's easiest white-box pavilion and its most chaotic competition image. Procurement depends on the middle of the workload.
Freeze the source view, pixel dimensions, brief, references, and acceptance checks. Give every candidate tool the closest equivalent input it supports. If one works directly inside Revit or SketchUp and another accepts only an exported image, record that difference instead of forcing false symmetry. Direct integration is a claimed differentiator in the Chaos comparison of six architectural AI renderers. Integration time belongs in the result because it is part of the route a studio is buying.
Run the test three times on different days or periods. Cloud conditions vary, and one lucky queue is not a capacity plan. Keep the fastest, median, and slowest completion time. Note failures, service interruptions, and credit consumption. Do not delete an abandoned generation from the ledger merely because it never reached the shortlist.
Count accepted images, not generated images
Throughput claims often use generations per hour. That denominator rewards noise. A tool producing 60 images with two acceptable results is not necessarily faster than one producing 12 images with six acceptable results. Calculate acceptance yield by dividing accepted outputs by total outputs, then calculate active minutes per accepted image.
Also report correction minutes per accepted image. This reveals two tools that reach the same final quality through different labor patterns. One may need careful setup and little repair. Another may produce a strong mood immediately but require local reconstruction around windows and edges. The better route depends on who is available and what the image must prove.
Keep the prompt rewrite ledger
Every changed adjective, reordered instruction, new reference, strength adjustment, crop, and seed search belongs in the record. Prompt work can look like harmless experimentation while consuming the largest active block. Save each submitted instruction with its result. If a vendor specialist supplies a successful prompt, time a normal team member reproducing it without coaching.
This is not punishment for iteration. Iteration is part of design. It is simply a refusal to classify design labor as machine speed.
Price the median route
Once the three runs are complete, use the median total time for planning. Multiply active minutes by the appropriate internal labor rate, add credits or subscription allocation, and note any finishing software used. Keep the slow run as a deadline-risk case. A procurement decision based only on the best run will fail exactly when the queue is busy and the client is waiting.
Compare the AI route with the studio's current route for the same stated use. Do not compare a concept image with a polished conventional render. Do not charge the old route for model preparation while giving the AI route a prepared screenshot for free. Start both clocks at the same project state and stop them at the same acceptance gate.
Our take: speed is a property of the system
The current tool comparisons are right to care about speed. A shorter feedback loop can change when visualization enters design and how many options a small team can inspect. But speed does not live inside the generator alone. It emerges from the model view, integration, brief, queue, output reliability, correction route, reviewer, and required use.
A 30-second generation is valuable evidence. It is one row in the ledger, not the headline number. Start the clock before upload. Stop it after approval.
Editorial basis: the 12 September 2026 ArchiGen AI intel sweep, including a community report of sub-minute SketchUp screenshot rendering and current tool comparisons emphasizing workflow speed. Product capabilities were checked against Chaos's June and September 2026 published guides. No hands-on product test is claimed.