AI panel and AI canvas

The AI panel edits a 2D drawing from one sentence. It shows a preview first; nothing changes until you apply, and applying is a single undo step. The AI panel is available in the two 2D workspaces; the 3D modeler, 2D drafting and vector design also have the AI canvas, which turns photos, plans and sketches into editable elements.

Where it is

• CAD drafting: type AI in the command line to open the AI panel (titled "AI assistant"); type AIIMAGE or AISKETCH to open the AI canvas.
• Vector design: menu "Window › AI" opens the AI panel; "Window › AI canvas (image / freehand)" opens the AI canvas.
• 3D modeler: the "AI canvas" button in the toolbar. There is no AI panel in 3D; to change a 3D design with words, use the "Instruction" tab of the AI canvas.

What you need

• Sign in with Google, or enter an invite code. Everything else works without signing in; only AI needs it.
• Each account or invite code has a monthly quota. You are told when it runs out.
• Read-only drawings (a "view only" share link) can preview but cannot apply.
• The first time, you have to accept the AI notice (see Consent).
• Whether AI is available depends on this site's server configuration. If AI is not set up, you see "AI is not set up on this server. Everything else still works." Besides the monthly quota there is also a site-wide daily limit.

How to use it

1. Pick a mode"Edit objects" adds, changes or deletes objects in the drawing. "Drawing rules" changes drawing-level rules, such as which layer is filled solid when cut or which layer uses thin lines.
2. Select first (optional)Select the objects you want handled; the AI works on those first. With nothing selected it considers the whole drawing.
3. Type the requestFor example "fill the selected walls black", "add a border on layer A1", "fill the wall layer when cut, thin lines for windows". Click "Preview".
4. Read the previewThe panel lists how many objects are added, changed and deleted, the affected extent, warnings and a list of differences. Hovering a list item highlights that object on the canvas.
5. Apply or cancel"Apply" writes the changes to the drawing. "Cancel" leaves it untouched. After applying, Ctrl+Z rolls everything back in one step. If you edited the drawing after previewing, it is re-checked before applying.

Writing a good request

• Inserting fields (CAD workspace): you can say “insert the project name field at the bottom right of the title block”, “insert today’s date” or “insert the typical storey height”. The AI only gets the list of available fields (three date formats, plus the names of the project information and global parameters when the project has a 3D design); it never sees the values and writes only a reference. Your browser reads the value from the 3D design and fills it in, and from then on it updates like a field inserted with FIELD (UPDATEFIELD, and when the drawing opens). If the source cannot be read, “----” is shown for now. The source design is always the local default; the AI cannot choose one. For sheet number and name fields, use FIELD in a layout.
• One thing at a time. "Fill the walls black" is more reliable than "fill the walls, thin lines for windows and add a border".
• Refer to objects by layer name, for example "the WALL layer". The AI sees the layer list.
• Select first to make the scope explicit.
• The AI does not see text content or the full geometry. "Change the text that says X" will not work.
• Read the warnings before applying.

AI canvas

The AI canvas turns a photo, a scanned plan, a PDF or a hand sketch into editable elements: walls, doors, windows, rooms and levels in 3D, and lines, text and dimensions on their own layers in 2D. In 3D you can also change a design with a sentence. Results are always shown as a preview first and are written to the design only when you press "Accept".

• Opening it: the "AI canvas" button in the 3D modeler toolbar; AIIMAGE (image) or AISKETCH (freehand) in the 2D drafting command line, or "AI canvas (image / freehand)…" in its application menu; "Window › AI canvas (image / freehand)" in vector design. It cannot be opened on a read-only design. The dialog has three tabs: "Image → elements", "Freehand" and "Instruction".
• Model: when the server uses Anthropic, the "Model" drop-down chooses Claude Sonnet 5.5 (default) or Claude Opus 5.5. Opus is slower and more careful and costs the same quota. Every model call counts toward this month's AI quota, including the automatic correction retries when the output fails the checks (up to 3 attempts); the number of attempts is shown below the preview. The choice is remembered in this browser. When the server uses Google Gemini (as the live site currently does), this drop-down is not shown and the server sets the model.
• While the dialog is open, workspace shortcuts do not fire. × closes it and cancels a running request.

Image → elements

1. Choose a filePNG, JPG or PDF, up to 25 MB. A PDF is converted one page at a time: enter the page in "PDF page"; if you change the page after choosing the file, that page is converted again about 0.4 seconds after you stop typing. The image is scaled down in the browser to 1568 pixels on the long side and converted to JPEG before it is sent.
2. Calibrate the scale (optional)Click "Calibrate: pick two points", click two points on the image, and type their real distance in metres in "Real distance (m)". The readout shows "Scale: 1 px = … cm". Calibration is optional: if the AI reads a dimension on the drawing it proposes a scale and asks you to confirm it; your own calibration takes priority.
3. Notes (optional)For example "walls 20 cm thick, 3 storeys of 3.2 m".
4. Place originX and Y in "Place origin (m)" are where the bottom-left corner of the image lands in the design.
5. Read with AIClick "Read with AI". The walls, doors, windows, dimensions and room names the AI found are drawn over the image, and the preview appears below. Changing the scale or origin afterwards recalculates the preview locally without calling the AI again.

• Result in 3D: elements go on the current plan-mode level, otherwise on the lowest level; a design without levels first gets 3 levels of 3.2 m. If the drawing or the notes give a number of storeys, levels are added above (not when the levels table is pinned). Walls use the built-in type with the closest thickness; doors and windows go on the nearest wall within 0.6 m, using the built-in type with the closest width, and ones that do not fit are skipped with a message. Dimensions are used only to check the scale; no dimensions are created in 3D. At most 1200 elements at a time: crop the image or work in parts.
• Result in 2D: missing layers are created automatically: A-WALL (walls; two lines when the thickness is known, otherwise a centreline), A-DOOR (door leaf and 90° swing), A-GLAZ (windows), A-AREA-IDEN (room names), A-ANNO-DIMS (aligned dimensions), A-ANNO-TEXT (text). Metres are converted to the drawing units. A locked or frozen target layer is refused; unlock it first.
• If a dimension written on the drawing differs by more than 10% from the length measured at the current scale, you get a warning: the scale or the reading may be wrong.

Survey sketches: dimensions rule

Hand-measured survey sketches are often not drawn to scale: the lines are only roughly placed, and the written dimensions are what is accurate. After loading the image, tick "Survey sketch: dimensions rule" (off by default). "Read with AI" then asks the AI to link every dimension number to the wall it measures, and to return only the text of numbers it cannot read instead of guessing. The wall positions are then fitted to those dimensions in the browser, without calling the AI again.

• Fitting rules: a dimension is the length of a wall, the distance between two parallel walls (centre lines or clear distance), or the distance along the dimension between the wall corners its ends point to; when the AI does not say which, it is worked out from the position. End points less than 0.3 m apart become one point; walls within 10° of horizontal or vertical are straightened; T-junctions and walls on one line stay on that line. The segments of a dimension chain must add up to the overall dimension, and an unreadable segment is worked out from the overall dimension.
• Scale: without a calibration, the median of "dimension value ÷ pixel length" is the starting value; the final geometry follows the dimensions.
• Summary and lists: shows "Constrained walls n/N" and the largest residual, and lists conflicting dimensions (more than 5 cm or 2% off after fitting, or a chain that does not add up), unreadable dimensions, dimensions not linked to a wall, and free walls (no dimension at all; they keep their sketched shape). Click an item to highlight it on the image; type a value (m) next to it to add a dimension or a wall length, and the fit is recalculated locally.
• Overlay: "Fitted" draws the fitted walls (dashed = free walls); "As drawn" draws the positions the AI read.
• Accept: creates the fitted geometry; doors and windows move with their walls, and rooms and text move with the fitted proportions. Limits, preview and the single undo step are the same as for a normal reading, in both 3D and 2D.
• At most 400 walls and 200 dimensions. If the fit takes longer than 2 seconds, does not converge or goes beyond ±10 km, the drawn geometry is used and the preview warnings say so.

Multi-floor batch

The "Multi-floor" tab handles a whole set of floor plans at once: each image goes to its own level, and the levels are aligned automatically.

1. Add imagesPick several PNG, JPG or PDF files at once; every PDF page counts as one image, up to 20 in total. Each image gets a row with a thumbnail, level, floor height, scale and reading status.
2. Set levels and heightsIn 3D the rows default to the existing levels in order (1F, 2F…). You can also choose "+ New level" and type a name. A new level is stacked on top of the previous image (its elevation plus its floor height); with no name it is numbered after the neighbouring levels. Wall height equals the row's floor height.
3. ScaleImage 1 uses "Calibrate: pick two points" or a dimension read by the AI. The other images default to "Same as image 1" (converted by the two images' resize ratios, which suits one scan set or one PDF); switch to "Calibrate separately" if needed. Click a thumbnail to enlarge that image and calibrate it.
4. Read all with AIEach image is read in turn by the same reading as "Image → elements". The tab shows the estimated number of calls (1 per image plus up to 2 automatic retries when validation fails); every call counts toward the quota. Cancelling keeps the finished images; pressing again reads only the unfinished ones.
5. AlignmentEach level's outline is the centreline of its outermost walls. It is matched to the reference level (image 1 by default) with a translation and a rotation (0/90/180/270°, plus up to ±5°); levels whose scale is not their own calibration are also scaled uniformly. Each row shows the mean deviation in cm and warns above 50 cm; after clicking the thumbnail you can enter X and Y (m) and an angle (°) under "Manual adjust". Alignment only moves each level as a whole; the geometry inside a level is unchanged. The overlay below shows each level's walls in its own colour.
6. Confirm and createTick "Confirm" per image or press "Confirm all", then "Preview creation". Accepting creates all confirmed levels together as one undo step.

• 3D: missing levels are created automatically (not when the levels table is pinned; pick existing levels instead). Two images cannot share a level. Batch totals are limited to 2,000 walls, 1,500 doors/windows and 800 rooms; split larger sets into batches.
• 2D: each level is drawn as its own plan, side by side from left to right, with a title such as "1F plan" below it; layers are the same as "Image → elements".
• Images stay only in the panel's memory and are discarded when it closes; each image sends the same data as a single reading.

Freehand

• Draw on the pad with a mouse, pen or finger. One grid cell is 1 m; "zoom" offers 40 × 24 m, 20 × 12 m (default) and 10 × 6 m. "Snap 10 cm" snaps points to 10 cm. There are "Undo stroke" and "Clear"; at most 200 strokes.
• Beautify (on by default): straightens strokes, snaps angles to multiples of 45°, and joins or closes endpoints within 0.3 m.
• Stroke classes: every stroke starts as "Wall". Shift+click a stroke to change its class: Wall, Door, Window, Furniture, Annotation, Ignore.
• Converting: "Convert (no AI)" converts locally, sends nothing and needs neither consent nor quota; "Classify with AI" lets the AI decide the class of each stroke and then converts.
• Each segment of a wall stroke becomes a wall. With "Classify with AI" the AI also estimates wall thickness: in 3D the wall gets the built-in wall type with the closest thickness; in 2D it is drawn as a double line. Without a thickness, 3D uses the exterior wall type and 2D draws a centreline. A door or window stroke places an opening at its centre, as wide as the longer side of the stroke's bounding box (0.6–3 m). Furniture and annotation strokes have no 3D element and are not converted; in 2D they become lines on the A-FURN and A-ANNO-SKCH layers.

Instruction (3D)

• On the "Instruction" tab, change the 3D design with a sentence, for example "add a 90 cm door between these walls" or "make it 12 floors of 3 m". Elements you select first are "these" in the instruction. Click "Preview" or press Ctrl+Enter (⌘+Enter on a Mac).
• The AI can only set parameters, add nodes, connect, remove, add or change walls, and add doors, windows, rooms and levels, up to 200 operations per batch. Pinned elements are never changed. The preview has a small plan: grey = unchanged, blue = new or changed, red dashed = removed. A batch that removes more than half of the nodes gets an extra warning.
• Analysis results and global parameters: the AI also sees the numeric results of the analysis nodes in the design (cost, area summary, regulation summary, structure summary, query, schedule summary…) (only numbers and yes/no, up to 50 values) and the list of global parameters (name, type, current value, range), so you can say “use 95% of the floor-area ratio” or “set the window width to the largest value in range”. The AI may change only constant global parameter values, and only when the parameter is not in the “Regulation” group, its value does not flow into a regulation node, it is not pinned, not driven by the graph and not a formula; the value must be within the minimum / maximum you set (with no range: lengths ±1000 m, angles ±360°); at most 20 per batch. Changes are listed in the preview as usual and are written only when you press “Accept”; Ctrl+Z undoes them in one step. The AI never connects analysis results back to elements or global parameters. Use the “Regulation” column of the Global Parameters dialog to mark or unmark a parameter; ones created by regulation features are marked automatically, and in older designs a group named like “Regulation” counts as marked.
• For text instructions in 2D use the AI panel (see How to use it); in 2D the "Instruction" tab has an "Open AI assistant" button.

Layout generation and design description

• Layout generation (the "Layout" tab of the 3D AI canvas, or the command ailayout): choose the outline (the mass floor plate of the current level, a floor on this level, or select a closed curve or a room first and press "Refresh"), fill in the use, the number of floors and the circulation preference, and list the rooms one per line as "name area [×count]", for example "Bedroom 12 ×3". Press "Generate with AI" to get 2–3 schemes.
• The AI schemes are checked first: every room must be a simple polygon inside the outline, rooms must not overlap, and the total area must not exceed the outline. Rooms slightly outside or overlapping are clipped by the program and marked "AI (repaired)"; if a scheme still fails, the AI is asked to fix it, and if it still fails a "deterministic split" is used instead and marked "not AI". If you do not want to use AI, press "No AI: split by area" and nothing is sent.
• The preview shows the scheme overlay, the achieved area rate of each requirement, the usage rate, the area counted toward FAR by use, and hints (a room with no exterior wall, a corridor or room that is too narrow, requirements that were not placed). The hints are not a code check; verify against the regulations yourself.
• Choose "Room separation lines + rooms" or "Walls + rooms" (edges on the outline get RC exterior walls, the others interior walls), then "Accept (one undo step)" creates them on the current level; Ctrl+Z undoes it in one step. Edges that already have a wall or separation line are skipped by default. Only the outline coordinates, the room list and the use names are sent.
• Design description (the "Design description" tab, or the command aidescribe): from the levels, the mass floor-plate area, the rooms summarised by use, the numeric results of analysis nodes (regulation summary, green building, environment, structure summary…) and the wall-type materials, it produces text in the style of a building-permit description (Traditional Chinese or English). "Design facts sent to the AI" can be expanded for inspection: only numbers and cleaned names, no project info, email or share links.
• The text can be edited directly (# title, ## heading). Numbers the AI wrote that have no source in the design data are listed for you to check. You can also press "No AI: template" to lay it out directly from the design data. When done, choose a sheet (or "New sheet") and press "Insert on sheet"; a drafting view is created and placed on it, as one undo step; or export Markdown or DOCX.

Preview and accept

• The preview lists the AI's explanation, the scale, the numbers of added, changed and removed items and any warnings, with the reminder "AI output can be wrong: compare with the image/design before accepting."
• If the AI output fails the format and scope checks, the AI is asked to fix it automatically, up to 2 more times, and each call counts as AI usage; if all 3 fail, the design is not changed.
• "Accept (one undo step)" writes the result; Ctrl+Z undoes it in one step. "Reject" discards the preview and leaves the design unchanged. If you changed the design after the preview, the result is re-applied on accept; if it no longer applies, preview again. If a command or operation is in progress, finish it or press Esc first. A read-only design can show the preview but cannot apply it.
• The AI can misread lines, scale and the positions of doors and windows. Compare with the original before accepting, and check again afterwards.

AI branches and diff

• Every AI preview creates an "AI branch" on this device: it keeps the proposed changes and a fingerprint of the main model at that moment, and does not write to the main model. Each design keeps its latest 5 branches, and they survive a page reload (listed at the bottom of the AI canvas and the AI assistant panel). Branches are not saved in the design file and are not uploaded to the cloud.
• Press "Diff (approve item by item)" to open the diff view. In 3D the viewport is coloured (added = green ghost, removed = red, modified = yellow) and you can switch between Before and After; in 2D the panel draws a diff in the same colours. The list shows each change with its element type, name and parameter values before and after, and every item has a tick box. When you tick an item, the items it needs (for example the new wall that hosts a door) are applied too.
• Below the 3D diff, the effect on the code checks after applying the ticked items is listed (BCR, FAR, height, §164 / envelope and so on). Checks that go from pass to fail are highlighted in red.
• "Apply ticked" writes to the main model as one undo step (one Ctrl+Z undoes it), and the same step adds an approval record to the design settings (who, when, a summary of the AI request, how many items were applied; up to 30 records; for 2D drawings the record is kept in the drawing file and syncs with saving, the cloud and collaboration). "Reject" discards the branch and leaves the main model unchanged. Accepting directly in the preview approves all items.
• Conflicts: if the main model was edited after the branch was created, every item is checked. Elements that were changed or deleted are marked as conflicts and cannot be applied; use "Apply non-conflicting only", or "Regenerate from the latest" to ask the AI again with the same request (3D instructions).
• Collaboration: while co-editing, branches stay on your device and nobody else sees them; changes sync to everyone like any other edit only after you approve them.
• The Multi-floor, Layout and Description tabs of the AI canvas work the same way: a multi-floor preview creates a branch automatically, and in Layout and Description "Diff (approve item by item)" creates a branch for the current scheme or text. Accepting or inserting directly also adds an approval record. Clicking a row in the list selects that element and moves the view to it (3D framing, 2D zoom). If an input source or a referenced element (for example the wall that hosts a door) changed after the branch was created, the row is marked "Upstream changed"; this does not block applying, but check the result.

Consent

• The first time you use the AI canvas, "Before using AI" is shown at the top; nothing is sent until you press "I understand, enable AI". The AI panel shows the same notice the first time you press "Preview". "Convert (no AI)" needs no consent.
• Consent is remembered in this browser only (in a private window, only until the page is closed). To withdraw it: Account menu › "Account" ("Sign in" when signed out), section "AI features", press "Withdraw AI acknowledgement"; the notice is shown again before you next use the AI canvas or AI assistant. Withdrawing does not affect requests already sent or results already accepted.
• The AI canvas sends the downscaled JPEG image, the coordinates of freehand strokes, the text you type and a summary of the design (for 3D instructions also the numeric results of analysis nodes and the names and values of global parameters) to the AI provider configured on the server (Google Gemini API or Anthropic) through this site's server. It does not send the design name, your account or email; email, phone and ID-number patterns in text are replaced first. Synhedra does not keep images, text or replies, the server logs metadata only, and the image is discarded when the panel closes. The live site currently uses the Google Gemini API; on Google's unpaid Gemini API tier, Google may use submitted content (including images and text) to improve its products, which may include human review; this does not apply on the paid tier. See the AI section of the privacy notice.

What data is sent

The panel says it too: "Only layers, selection and a statistics summary are sent, never the whole drawing or personal data." In detail:

• Sent: your request text, the layer list, the current selection, the drawing extent, counts per entity type, and a limited number of entity samples (handle, type, layer, bounding box). Large drawings send fewer samples.
• Not sent: the full geometry, the content of text and multiline text, file names, design names, your account data or email. Layer names and labels are scrubbed of email, phone and ID-number patterns first.
• Route: the request goes to this site's server, which forwards it to the language-model service (Google Gemini API or Anthropic Claude, depending on the server configuration; Anthropic when both are configured; the live site currently uses the Google Gemini API). The model API key lives only on the server; the browser never sees it.
• What comes back: the model's proposed edits are checked for format and scope before they become the preview.
• Logs: the server records usage counts and timing metadata only, not your request text. See the privacy notice.

Messages