Image to CAD · Methodology
How it works, and what its numbers actually mean
This page exists so you can judge the output rather than trust it. It describes the pipeline, the measurements behind every figure we publish, and the cases where the tool does not work.
What it does
It turns architectural imagery into editable vector geometry — DXF, SVG and PDF. Drawings, scans and sketches are vectorised entirely in your browser and never leave your device. Photographs take one extra step: they are drafted into clean architectural linework by an image model in the cloud, and that linework is then vectorised locally into CAD entities with layers, true arcs and shared junctions.
What “elevation” means here
When you photograph a building, the drafting stage is instructed to produce a true orthographic elevation — floor lines exactly horizontal, wall edges exactly vertical, nothing converging. Atlasly then checks that result against a fixed geometric standard: both line families must sit within 2.2° of their axis, with enough linework in each family to measure. Drawings that miss it are not offered as elevations.
Every line is separately checked against your original photograph. Lines with no supporting evidence in the photo are marked as inferred and drawn dashed, so you can see which geometry the system is confident about and which it is not.
What it does not mean
- It is not a measured survey. The elevation is squared by drafting, not by surveying. Nothing on the drawing was measured on site.
- It is not photogrammetry. No 3D reconstruction takes place, and no camera position is solved from multiple views.
- It is not construction-certified geometry. Verify dimensions before anyone builds, prices or submits from them.
- Passing the check is not proof of correctness. It means the drawing meets a geometric standard and its linework is backed by the photograph — not that every detail is right.
What scale means
Exports are unitless until you calibrate. Choose two points a known distance apart and enter that distance, and exports then carry real units. Calibration scales the drawing; it does not turn a photograph into a survey, and it cannot correct geometry that was never accurate.
Where a scanned drawing carries a printed dimension, the text is read and offered as a scale suggestion — always for you to confirm. It is never applied silently, because our own testing caught the reader misreading 136′-2″ as 156′-2″.
How reliable it is
The drafting stage is generative, which means the same photograph can produce slightly different drawings on different attempts. That is measurable, so we measure it rather than quoting a single lucky run: a fixed set of real camera photographs is generated repeatedly and scored by the same standard each time.
- 95%
- Single attempt clears the elevation gate
- 61 of 64 generations across a stratified 8-image slice (E35)
- 99.5%
- Within one automatic retry
- retry fires only when the first attempt misses the gate (E36)
- 7 of 8
- Images that pass reliably on their own
- one image is variable: 6 of 8 attempts (E35)
- 1.05
- Average generations per result
- retries are rare, so cost is essentially unchanged
These come from a slice deliberately weighted toward difficult photographs — oblique, ultra-wide, night and heavily vegetated — so they are not a flattering average. With eight repeats per image the confidence intervals are wide; an image measured at 8 of 8 is reliably good, but the true rate could still be as low as roughly two in three.
Where it fails
Occlusion is the real limit
If a tree, scaffold or vehicle hides part of the wall, that geometry is not in your photograph and no drawing method can recover it. Our worst measured case is a cedar across a façade: the drawing is highly faithful to the photo (99% of its linework is backed by visible structure) and still fails as an elevation, because what it faithfully draws is a tree.
Vegetation can still fool the quality check
Bare winter branches are long and straight, and our geometric check can score them as if they were architecture. We tried to separate them automatically using image texture and the attempt failed — it also flagged a parked truck as architecture, because chrome and panel seams are just as strongly oriented as masonry. That result is published rather than hidden, and the fix needs real semantic segmentation.
Curtain walls do not become windows
On fully glazed towers our window detector boxes the entire building as one opening. Rather than label that as windows, the system rejects it: openings stay unlabelled instead of wrong.
Repeated openings are conservative
A confirmed opening only becomes a reusable CAD block when its strokes belong to nothing else. A window head is frequently also a string course, and collapsing shared linework would silently delete real geometry. On a typical terrace that means a minority of confirmed openings become blocks and the rest stay as labelled linework.
Your images
Drawings, scans and sketches are processed entirely in your browser and are never uploaded. Photographs are sent to our cloud drafting step, and the page tells you so before you upload. Your work is not used as training data.
Checking our work
The measurements above come from a benchmark suite kept in the repository alongside a research log that records every experiment, including the ones that failed and were rejected. Exported files are validated by an independent DXF library rather than by the code that wrote them.
Back to the converter