Take one plain exterior view with a difficult corner, repetitive windows, a recessed entrance, and two specified materials. Run it through a prompt-first generator, a BIM-connected renderer, a hosted image-to-image service, and a ComfyUI graph with structural conditioning. Do not begin by choosing the prettiest output. Mark what each system changed without permission.
The pattern of unauthorized changes is more useful than a composite quality score. It tells you what evidence the system reads, what it treats as optional, and which correction control the workflow will need in production.
Failure signatures are evidence, not bad luck
Generative systems create plausible images from different combinations of text, pixels, geometry, depth, edges, masks, and reference inputs. When one source of evidence is weak, the system fills the gap with a likely visual answer. The answer may be attractive and still be wrong for the project.
A prompt-first system has strong freedom and weak obligation to the source building unless reference controls carry structure. An image-to-image tool sees the source pixels but may treat local geometry as negotiable when denoise or creativity rises. A BIM-connected tool begins closer to project geometry but may simplify materials or populate details according to its presets. A node graph can preserve several kinds of structure, provided the operator builds and tests those controls correctly.
Do not ask whether the AI made a mistake. Ask which project fact had no strong path into the result.
Signature one: silhouette and opening drift
Look first at the roofline, corners, slab edges, major recesses, and window count. If the mass becomes taller, bays merge, or an entrance moves, the workflow is weak at global or mid-scale structure. More descriptive prose rarely fixes this reliably because the missing information is spatial.
For a BIM-connected tool, reduce geometry override or creativity and test whether a closer model view improves the result. For image-to-image generation, lower denoise and add depth or edge conditioning. In ComfyUI, compare a depth pass with a line or edge pass: depth protects spatial order, while edges often defend openings and crisp boundaries more directly.
If structural controls make the image dull before they make it accurate, separate the task. Produce a faithful base first, then apply bounded material and atmosphere changes. A system that needs permission to redesign the building before it can improve the light is a poor fit for late-stage work.
Signature two: material identity substitution
The geometry survives, but limestone becomes concrete, standing-seam metal becomes timber, or clear glass turns mirrored blue. This signature shows that shape evidence is stronger than semantic material evidence. Source pixels alone may not carry a specification, especially in a clay render or a low-resolution viewport.
Give each protected material a name, location, and observable behavior. “Warm stone” is weak. “Honed pale limestone on the two solid end walls, low reflectance, fine horizontal joints” can be checked. Use region masks or material IDs when available. Keep one reference per material role instead of blending several interiors or facades into a general style cue.
Watch junctions. A generator may reproduce the color of brick while erasing its coursing at corners, or keep timber tone while stretching grain across separate boards. Material fidelity includes scale, direction, edge behavior, and adjacency, not just hue.
Signature three: local detail invention
Door handles multiply, railings lose supports, paving joints terminate oddly, and furniture develops extra legs. These errors appear where the source has small, ambiguous patterns. Upscaling can sharpen them without making them correct.
Decide whether the detail is protected, replaceable, or too small to judge. Protected details need a stronger source crop, mask, edge input, or ordinary compositing. Replaceable details can be regenerated locally. Tiny elements should not be approved from a reduced preview; inspect at delivery resolution.
A useful test is repeat generation with the same source and changed seed. If a detail changes identity every time, the system does not know it. Do not select one lucky version and call the element controlled. Either provide more evidence or remove that feature from the AI pass.
Signature four: view-set inconsistency
A single image looks convincing, but the next camera shows different cladding, planting, furniture, weather, or facade proportions. This is the signature most beauty-shot comparisons ignore. Each independent generation has rebuilt the project from partial evidence.
Test three views from the start. Include one shared corner or repeated room element so differences become obvious. Hold prompts, presets, model versions, and material references steady. Where the system exposes seeds or reusable style settings, record them, but do not assume a shared seed guarantees architectural consistency across different compositions.
For fixed objects, use the 3D scene or conventional render as the common source of truth. Reserve generation for bounded regions that can vary without contradicting another view, such as sky, distant context, or loose planting. If every camera needs extensive independent repainting, budget the result as illustration rather than synchronized visualization.
| Failure signature | Likely missing control | First response |
|---|---|---|
| Moved openings | Geometry or edge evidence | Lower change strength, add depth or edges |
| Wrong material | Material identity and region | Name, mask, and reference the surface |
| Invented small parts | Local source detail | Crop, mask, composite, or omit |
| Different views | Shared project state | Test a set and anchor fixed content in 3D |
| Style overwhelms design | Reference-role limits | Separate appearance from protected facts |
Signature five: correction spill
Ask to replace one tree and the tool changes the facade color, clouds, cars, and glazing. Correction spill means the editing boundary is weak. A chat instruction may describe the desired change but still leave the system free to reinterpret the entire frame.
Use a tight mask with enough feathering to blend but not enough to cross protected edges. Lower generation strength. Keep the accepted image as the source, not an earlier candidate. Compare the result outside the mask as well as inside it. A correction fails if untouched pixels move in a way the reviewer would notice.
Some hosted tools make local editing easy but hide the exact model or settings. Some node workflows make every parameter available but demand more setup. Judge both by the smallest reliable correction boundary, not the number of controls shown in the interface.
Build a one-page failure log
For each trial, save the source, result, settings, and a marked overlay. Classify each defect as silhouette, opening, material, junction, detail, content, atmosphere, text, or cross-view inconsistency. Record whether the defect was fixed by a prompt, parameter, structural input, mask, composite, or complete rerun.
After ten images, the log exposes the true workflow. A tool may generate excellent mood but repeatedly move geometry. Another may preserve the model while needing material finishing. The first belongs earlier in design or inside bounded ideation. The second may be the better production base.
Track recurrence, not just severity. One dramatic defect is easy to notice. A minor mullion error that returns in every camera consumes more review time and carries greater issue risk. The best fit is the system whose predictable errors your team can catch and correct cheaply.
Our take
The many 2026 renderer rankings are trying to turn unlike workflows into one ordered list. Architects need a different result: a map from project facts to controls, followed by a record of what escapes those controls.
Circle the first unauthorized window. That mark knows more than the leaderboard.
Written from the 28 August 2026 intel sweep, which surfaced current tool comparisons and community debate about Midjourney freedom, BIM-connected rendering, and the setup cost of ComfyUI plus ControlNet. ArchiGen AI carries no sponsored placements.