Three new 2026 lists reached today's search sweep with different winners. Chaos compares six products and emphasizes BIM integration. Gendo ranks ten products and names its own collaborative design canvas best overall. Syntina compares eight and organizes the choice around practical fit. The disagreement is useful. It exposes the variable most ranking tables compress: how project information enters the tool and how an approved result gets back to the team.
A tool that accepts a native model is not automatically better than one that accepts pixels. It is solving a different transfer problem. The mistake is awarding both an “input support” checkmark without recording what crosses the boundary, what gets flattened, and what must be rebuilt after each revision.
The useful unit of comparison is not the generated image. It is the round trip from project source to reviewed decision.
Start with the artifact you actually have
Write down the source before opening a pricing page. It may be a coordinated Revit model, a Rhino study, a SketchUp scene, a clay viewport export, a floor plan, a loose sketch, a reference photograph, or only a sentence. Those sources carry different amounts of design truth.
A native connection can preserve the camera and provide a faster capture loop. An uploaded image can travel between many tools but freezes model structure into pixels. A text prompt starts with almost no project geometry. A floor plan contains literal spatial evidence but little information about materials, height, or light. Calling all four “inputs” is accurate and nearly useless.
For each candidate, record the exact accepted artifact, file type, maximum dimensions, host application and version, cloud transfer, and preprocessing step. If the source must become a JPEG, note who makes that export and which layers, annotations, crop settings, and color transforms survive it.
Give every tool a seven-column route card
| Column | Question | Evidence |
|---|---|---|
| Source | What leaves the project? | Named file or host view |
| Transfer | Plugin, upload, paste, or prompt? | Recorded steps |
| Control | What constrains geometry? | Setting, mask, map, or none |
| Output | What comes back? | Image, video, or model |
| Revision | Can one fact change locally? | Measured repair attempt |
| Return | Where is approval recorded? | Project folder and owner |
| Loss | What must be reconstructed? | Explicit loss list |
This card prevents a familiar comparison error. Chaos lists direct connections for Veras and describes Midjourney as having no BIM integration. That distinction matters, but it still does not measure how accurately one particular source survives a generation. Integration is a route property, not a quality verdict.
The same caution applies to a web tool that accepts sketches, screenshots, or plans. Breadth of accepted formats can shorten intake while leaving the operator to judge what the model inferred. Ask which facts are read from the source, which are suggested by a prompt, and which appear because the generator found them visually plausible.
Count the conversion tax
Every boundary creates labor. A screenshot may need a clean view template, hidden categories, exposure correction, a crop, and naming. An upload may require a confidentiality check. A generated image may need download, renaming, color correction, redlining, and manual placement on a sheet. A second design revision can repeat the entire chain.
Measure the first route with a stopwatch, then measure a small revision. Use one controlled request, such as changing a cladding zone while retaining openings, camera, planting, and time of day. Record operator minutes, wait time, attempts, credits, and changes outside the target.
A fast first image can carry a large conversion tax on revision two. A slower setup can become cheaper when the source view stays connected. Neither result should be assumed from the category. Test the route the studio expects to repeat.
Separate geometry control from geometry evidence
Ranking pages often use phrases such as geometry control, model fidelity, or architectural accuracy. Turn each phrase into a visible mechanism. Does the tool receive a native scene, a depth map, an edge map, an image, or only words? Is there a strength control? Can the operator select a region? Can the result be compared against the source with an overlay?
A control slider is evidence that the operator can request a different balance. It is not evidence that the output retained the correct bay count. A direct plugin proves the path exists. It does not prove every category, material, crop, or thin edge crossed correctly. A favorable sample proves one image was produced. It does not establish a repeatable rate.
Choose five countable facts in the source: floors, bays, doors, railings, and repeated lights work well. Add one material boundary and one contextual edge. Score the first output and the best of a fixed number of attempts. Keep the rejects. The result is modest, local evidence, which is more useful than a universal star rating.
Test the return path before buying
Most lists stop when the image appears. Practice work continues. The image needs a filename, project location, author, source reference, prompt or settings record, usage rights check, review note, and revision status. Someone must know whether it is exploratory, client-facing, or approved only as a mood reference.
Ask a second team member to find the source and repeat one change without a private explanation from the first operator. If they cannot identify the model view, settings, selected output, and approval note, the route depends on memory. That is a handoff failure even if the picture is excellent.
Also test exit. Can the team export original-resolution files and retain a usable record if the subscription ends? Can an administrator remove access? Can a project be deleted? Does a shared link reveal other work? These questions rarely move a public ranking, but they decide whether a route belongs in a real office.
Read vendor rankings as positioning
A vendor-authored comparison can still be informative. Chaos clearly states its criteria and places integration first. Gendo explains its product as a collaborative canvas and ranks it best overall. Those choices reveal how each company defines the problem and where its product wants to sit.
Use the page to build a shortlist and identify claims for checking. Do not treat the publisher's preferred category as your brief. Record ownership, commercial relationship, publication date, tested version, named inputs, number of trials, rejected outputs, pricing date, and whether the method can be reproduced. An absent disclosure is not proof of bad testing. It is missing evidence.
Then visit official product documentation for current compatibility, formats, limits, and terms. Pricing copied into a roundup can expire. A plugin list can gain or lose hosts. An engine can change while the product name stays fixed. Date every route card.
Our take: pick the route, then the renderer
The studio does not need one champion for every image. It may need a native-model route for controlled design options, an image route for finishing a chosen view, and a prompt-first route for references that make no claim about the current building.
Approve each tool for a named source, stage, output, and revision type. Keep a short loss list beside it. Retest when the host, engine, or team process changes.
If the route is invisible, the ranking is unfinished.
Editorial basis: the 30 September 2026 ArchiGen AI intel sweep and the published comparison methods in Chaos's six-tool guide, Gendo's ten-tool ranking, and Syntina's eight-tool comparison. Publisher identity and stated criteria were checked on 30 September 2026. Vendor pages establish their own claims and positioning, not independent performance. This article reports no hands-on test. The route card and test protocol are editorial recommendations.