The stopwatch starts before upload.

This morning's tool sweep produced a familiar table column: “main inputs.” It is useful to know whether a product accepts text, an image, a sketch or a host-model view. It is less useful to pretend those nouns arrive ready.

A Revit view may need categories hidden, entourage removed and a camera rebuilt. A screenshot may require a clean material pass. A ComfyUI graph may need depth and edge controls exported at matching dimensions. “Upload image” can conceal half an hour of judgement behind two tidy words.

That hidden work distorts comparisons. A ten-second generation fed by a carefully prepared source is not a ten-second workflow.

What belongs in a preparation ledger?

The ledger begins with the source project and ends when the renderer can accept the first honest run. It excludes generation, selection and repair, which belong in an intervention ledger. Keeping the boundary clear stops setup labor from disappearing into “normal project work.”

Use staff minutes, not elapsed computer time alone. A base render may occupy the machine for eight minutes while the operator works elsewhere. A category-cleanup pass may take eight fully attentive minutes from somebody who knows the model. Those minutes have different costs and should not be mashed into one duration.

“BIM input” is a starting claim

A native plugin can reduce transfer work. It can read the active view without a PNG round trip, inherit a camera and expose model geometry directly. That is real value. It does not prove the active view is fit for generation.

Test what must be changed before capture. Can temporary elements be excluded without altering documentation views? Does the plugin respect section boxes, hidden categories and linked files? Must materials be simplified so the model does not send contradictory cues? ArchiGen's guide to BIM integration in AI renderers separates the badge from the data that actually survives.

The fair comparison is not plugin versus upload as abstract categories. It is the preparation needed to get the same project evidence into each route. A plugin may save export time but demand a special view template. A browser tool may accept the renderer's existing beauty pass with no project changes. Measure both.

Separate reusable setup from view-by-view tax

Some preparation pays rent across a whole project. Build one export preset, one clean view template and one reference register, then reuse them. Other tasks return for every camera: fixing foreground obstructions, regenerating a depth pass or drawing a local mask.

Record those buckets separately. A pipeline with 45 minutes of reusable setup and five minutes per view beats a pipeline with two minutes of setup and 15 minutes per view after the fourth view. For a single competition image, the answer reverses.

This is why a one-image demo settles very little. It rewards the route with the smallest opening ceremony, even if every subsequent view requires hand work. Run a set. Our cross-view consistency test explains why a client sees the collection, not the lucky frame.

Count the failed preparation too

A discarded source still consumed labor. If the first export used the wrong crop, the first depth map crushed the foreground or the first reference pushed the material in an unusable direction, keep those minutes in the record. Removing failed setup from the total gives the winning route a fictional past.

Label the cause. Operator error suggests training or a saved preset. Tool incompatibility suggests a recurring tax. Project-specific difficulty may not transfer at all. The point is not to punish experimentation. It is to show a practice how much experimentation sits between an ordinary project file and the clean input shown in a tutorial.

Do references count as input labor?

Yes. “Supports reference images” sounds like a control feature, and often is. The image still has to be found, cleared for use and made relevant to the decision.

A material reference cropped from a manufacturer sheet is different from a mood image pulled from an unknown portfolio. One carries product information. The other carries authorship and licensing questions. Both require somebody to decide what the model should copy and what it should ignore.

Log the search minutes and source. Add a one-line instruction for each reference: brick bond only, warm overcast light, planting density, chair silhouette. ArchiGen's reference provenance register makes that evidence portable instead of leaving it in a downloads folder called final-final.

Control passes are not free because the renderer made them

Depth, Canny, normals, Cryptomatte and object masks can hold geometry or make corrections addressable. They also have production conditions. Resolution must match. Near and far values need useful separation. Transparent and displaced objects can behave badly. A pass exported from the wrong camera is decorative paperwork.

Record whether the source application already produces the pass, whether an AI preprocessor derives it, and who checks it. Derived depth from a flat image may be convenient, but it is an estimate. A native depth pass has access to the scene, though it can still encode glass, vegetation and distant context poorly.

The native depth pass workflow is useful precisely because it treats preparation as a controlled deliverable. The gray image is not an incidental file. It is structural evidence.

How should a comparison publish the result?

Put preparation beside generation and correction, not in a footnote. Name the source state and operator skill. State the number of views. Show what was reusable.

Do not turn the last row into a universal score. Project state changes it. A clean BIM model, a polished beauty render and a napkin sketch impose different preparation. Publish the condition so another practice can recognize its own work.

ArchiGen's take: input is made, not found

Tool tables need an input column. Keep it. Then add the work required to produce that input at a standard the tool can use.

This correction is unfashionable because it makes instant rendering look less instant. Good. Architectural images have always rested on upstream decisions. AI changes where some of the labor sits; it does not repeal the labor.

Start the stopwatch when the project opens. The upload button has had enough free publicity.

Evidence note: the 24 October 2026 sweep surfaced Syntina's 19-tool comparison, which lists main inputs, and Chaos's six-tool comparison, which describes BIM integration as a comparison criterion. This article proposes ArchiGen's preparation-ledger method; it does not report a hands-on benchmark or rank named products.