Should architects copy a Veras source description into the prompt?
No. Use the generated description to audit what the model sees in the source image. Mark correct observations, missing design facts and mistaken labels, then write a prompt that separates fixed geometry from desired change. A caption reports appearances; an architectural brief assigns intent.
Veras looks at a concrete school extension and writes a competent caption. It sees the timber fins, children in a courtyard and warm evening light. It does not see that the fins are the planning condition everyone spent six months negotiating.
Copy the caption unchanged and the omission becomes policy. The renderer received a description of the picture, not a record of what must survive.
What did Veras 4.5 add?
The official Veras 4.5 release post moved Describe Source Image into the main panel. Chaos says Florence 2 generates short, medium or long descriptions, which users can copy as a prompt starting point. Changing the source clears the previous description, a sensible guard against applying yesterday's caption to today's image.
The feature is practical. A blank prompt box is surprisingly expensive when a view contains material, atmosphere, entourage and camera clues. A machine caption can supply nouns quickly and expose how the system parses the source before credits are spent on generation.
“Starting point” is doing serious work, though. A caption model is trained to say what appears in an image. It is not reading the brief, specification, design history or approval record. It cannot know which visible facts are essential and which are merely provisional.
Why is a good caption still a weak brief?
A caption compresses. That is its job. Ten windows become “a glazed facade.” Two brick tones become “red masonry.” A temporary hoarding may become a permanent fence because pixels do not arrive with issue sheets.
Architecture depends on distinctions that may look trivial. A retained wall and a new wall can share a finish. A protected tree and replaceable planting can occupy the same green patch. A caption can be visually correct and operationally wrong.
The problem sharpens when the prose sounds confident. Detailed descriptions create an impression of coverage, yet additional adjectives do not restore missing project knowledge. A long caption may name the weather, camera angle and mood while overlooking the one datum the edit must preserve.
Our reference-image brief assigns each image a specific job. Describe Source Image can help inventory what is visible. It cannot decide what authority the image should have.
How should the caption be audited?
Put the source beside the generated text and mark every phrase with one of four labels. This is a short review, not a literary exercise.
| Label | Meaning | Prompt action |
|---|---|---|
| Observed | Visibly accurate and relevant | Keep if it helps the requested edit |
| Inferred | Plausible but not confirmed by pixels | Verify or remove |
| Missing | A fixed design fact was omitted | Add as an explicit preservation rule |
| Misnamed | The caption chose the wrong object or material | Correct with project terminology |
Run short and long descriptions on the same source during a trial. The difference reveals what the model regards as optional detail. If a facade rhythm appears only in the long version, that rhythm needs human promotion into the fixed brief. If neither version notices a recessed entrance, name it yourself.
Keep the original description with the source and final prompt. The gap between them is a compact record of human judgment, which is more useful than a polished prompt with no visible ancestry.
What belongs in the rewritten prompt?
Separate facts from requests. Begin with the source type and scope: an exterior render of a named view, an interior viewport, a scanned sketch. State what must remain: massing, opening positions, camera, key material divisions and any approval-sensitive element. Then state the change.
“Warm modern school with timber fins” is a caption. “Keep the existing camera, two-storey massing, opening grid and full depth of the approved vertical fins; change only the grey rainscreen to dark red terracotta under neutral afternoon light” is an edit instruction.
The distinction does more than improve prompting. It creates a review test. Geometry can be compared against the fixed clause. Finish and light can be compared against the requested clause. The prompt has become a tiny acceptance document.
Use the change-budget method when the source is moving into ComfyUI. The same logic applies in Veras: list what may change, what must hold and which elements need a mask rather than eloquent pleading.
Can the caption reveal model bias?
It can reveal a local symptom. If the description calls board-marked concrete “stone,” reads a civic foyer as a hotel lobby or treats a planted roof as ground, note the mismatch before generation. That does not diagnose the whole model. It tells you which terms need correction in this source.
Watch what gets described first. Image captions often prioritize prominent objects and familiar scene categories. Architecture may be reduced to style while cars, people and weather receive lavish attention. That ordering matters if the copied text becomes a generation prompt, because repetition and specificity can steer emphasis.
A caption can also expose ambiguity in the image itself. If three attempts disagree about whether a wall is timber or bronze, the presentation may not communicate the finish as clearly as its author believes. Sometimes the machine is confused because the picture is coy.
Which description length should architects use?
Use short for a fast category check and long for omission hunting. Medium may be convenient for routine work, but length is not a quality score. Pick the setting that supports the review question.
For an early sketch, short may be enough to confirm that the system sees a courtyard rather than a roof terrace. For a developed exterior, long is more likely to surface secondary materials, entourage and lighting assumptions that could contaminate an edit. Compare, annotate, then discard whatever has no bearing on the task.
Do not keep a sentence because Florence 2 wrote it. Keep it because a reviewer can point to the source or the brief and defend it.
How does this fit a revision record?
Save four items together: source image, generated description, edited prompt and chosen output. Add the Veras build, engine and date. If the output drifts, the record shows whether the source was misread, the human correction was absent or the renderer ignored a clear constraint.
Our visual decision log provides the wider project record. The description audit fits inside it as evidence of interpretation, not evidence that the machine understood the design.
Our take: the useful part is the disagreement
Describe Source Image is strongest when treated as a mirror held up to the model. It shows what appears obvious to the captioner and, by omission, what remains private knowledge inside the project team.
Copying the text closes that inquiry too early. Correct it. Narrow it. Put the planning condition back in.
The caption gets the nouns. The architect assigns the stakes.
Evidence note: this is a proposed review method, not an ArchiGen hands-on benchmark. Veras feature claims were checked against the official 4.5 release post on 19 October 2026. The post identifies Florence 2, three description lengths, clipboard copying, cancellation and automatic clearing when the source changes. The workflow recommendations are ArchiGen's analysis.