Image Context

The Image Context viewer lets Supervertaler attach figure images to AI translation requests so the model can see the document’s drawings, photos, schematics or labelled parts in addition to the text. When a segment contains Figure 1, Fig. 2A, Table 3, etc., the matching image is shipped with the request so the AI’s translation is anchored in what the figure actually depicts — not just the words around it.

The same viewer doubles as Supervertaler’s DOCX image extractor: point it at a Word document and it pulls every embedded image out into a numbered folder of PNG files, ready to load straight back as AI context. So the most common flow is one button-press: extract from a DOCX → context is auto-loaded → start translating.

Where to find it

The viewer is folded into the Prompt Manager (since v1.10.176; it used to be a standalone AI sub-tab):

1. Switch to the ✨ AI tab → Prompt Manager sub-tab.
2. On the left of the panel, find Section 4 — Image Context.
3. Click the green Open ▸ button. The right-hand panel swaps from the Prompt Editor to the Image Context viewer.
4. The viewer’s ← Back to Prompt Editor button (or clicking any prompt in Section 5) returns you to the Prompt Editor.

Extract images from a DOCX

The viewer is a single toolbar at the top, with a results list + image preview below.

1. Add input files to the extraction queue:
   • 📄 Add DOCX — pick one DOCX file
   • 📁 Add Folder — add every DOCX file in a folder (batch input)
2. Choose an output strategy:
   • Enable Auto-folder to create an Images folder next to each DOCX
   • Or set a single output directory in the Output directory field
3. Set Prefix (default Fig.).
4. Click 🖼️ Extract Images.
5. The freshly-extracted folder is automatically loaded as AI context for figure-aware translation — no second click required.
6. Use 📂 Extracted Files (click to preview) below to preview images.

Filename detection

Since v1.10.190, extracted files are named after the caption visible in the document rather than sequential Fig. N.png numbers. So if your document labels figures as FIG. 7, the extracted file is FIG. 7.png — matching what the reader sees.

The detector recognises the following label vocabularies (case-insensitive):

Pattern	Example filename
FIG. N / FIGS. N	FIG. 7.png (patent figures)
Figure N / Figs N	Figure 7.png (academic / general)
Fig. N	Fig. 7.png
Table N	Table 3.png
Diagram N	Diagram 4.png
Chart N	Chart 2.png
Photo N / Photograph N	Photograph 12.png
Scheme N (chemistry)	Scheme 1.png
Plate N / Plate IV	Plate IV.png (Roman numerals OK)
Exhibit A (legal)	Exhibit A.png

The ID portion accepts N, Na, NB shapes (e.g. FIG. 6a, Table 3B). Documents that use Word’s built-in Caption paragraph style are detected even when no pattern matches — the leading sentence of the caption becomes the filename (capped at 80 chars). Images for which no caption can be detected fall back to the sequential Fig. N.png form.

🤖 AI label detection (opt-in)

For documents that don’t follow the standard label vocabularies — marketing copy, blog posts, cookbooks, photo essays, foreign-language documents — tick ”🤖 AI label” in the toolbar before clicking Extract Images. Each image that the text-pattern detector couldn’t label is sent to the active vision AI alongside surrounding text from the document, with a request to identify the figure’s label.

Requirements:

• A vision-capable model configured in Settings (Claude Sonnet/Opus 4.x, GPT-4o or newer, Gemini)
• Internet connection

Cost: roughly $0.005–$0.02 per AI-labelled image depending on provider — only spent on images the free text-pattern detector missed. A confirmation dialog showing the estimated extra cost appears before any API call is made; click Cancel to abort or run text-pattern only.

Pre-flight gates that may interrupt the run:

• No chat backend / API keys configured → friendly message, AI step skipped
• Active model doesn’t support vision → choose between proceeding text-only or cancelling
• Zero images need AI (all already pattern-detected) → silently skipped, no dialog, no cost

If the AI replies unlabelled for a given image, that image falls back to the sequential Fig. N.png form. So the AI step is purely additive — it can only improve filenames, never make them worse.

Note

Documents with conventional captions (patents, academic papers, technical specifications) don’t need this feature — the free text-pattern detector handles them perfectly. The AI fallback is for documents where captions follow no recognisable pattern.

Load a pre-existing folder of images

If you already have a folder of images ready (e.g. one you extracted in an earlier session, or one you assembled by hand):

1. Click 📁 Load Folder (the green button, to the right of Extract Images).
2. Pick the folder.
3. The images load into the AI context and populate the preview list below — just like a fresh extraction.

Note

“Add Folder” vs “Load Folder” — these are the two confusing buttons. Add Folder queues a folder of Word documents to extract images from. Load Folder loads a folder that already contains image files directly for AI context. The on-button tooltips spell out the distinction.

Filename → figure-reference matching

Supervertaler infers the figure reference from each image’s filename. Recognised patterns:

Filename example	Matched reference
Figure 1.png	1
Fig. 2A.jpg	2a
figure3-b.png	3b
Fig 10.tif	10

When the AI later sees Figure 1 or Fig. 2A in a segment’s source text, the matching image (case-insensitive, whitespace-/dash-/dot-normalised) is attached to the request. References that don’t match any loaded image are simply ignored — no error, no warning, the segment translates text-only.

How loaded images reach the AI

When AI translation runs (single-segment or batch), Supervertaler scans each segment’s source text for figure references. If a match is found AND a corresponding image file is loaded AND the active model supports vision (Claude Sonnet/Opus 4.x, GPT-4o or newer, Gemini), the image is base64-encoded (or passed as PIL data for Gemini) and attached to the request. Segments without figure references are translated text-only.

A line appears in the log for every match, e.g.

🖼️ Detected figure references in segment #42: 1, 2a
✅ Including 2 figure images: 1, 2a

If the model is text-only (older GPT, Ollama models without vision, etc.), the figures are silently skipped and a warning goes to the log so you know why visual grounding didn’t fire.

Using images with AutoPrompt (opt-in)

Since v1.10.178, the loaded figures can ALSO be sent to the AutoPrompt generator — not just to the per-segment translator. Section 2 of the Prompt Manager has a companion checkbox under the ✨ AutoPrompt button labelled “🖼️ Include loaded figure images”.

Tick it before clicking AutoPrompt to ship the loaded figures alongside the meta-prompt. The LLM then uses the drawings to lock terminology decisions directly into the generated translation prompt — “part 7 in Figure 1 is labelled ‘cylindrical sleeve’ → lock ‘mantelbuis’ → ‘cylindrical sleeve’ in the termbase” — instead of having to guess from textual references alone.

• Off by default — opting in is a deliberate per-project choice.
• Adds a small extra cost — roughly $0.05–$0.30 for 10–20 figures with a Sonnet-class model, more with Opus. Negligible vs. the value of a typical project (a €1000 patent will spend less than 0.1% on visual grounding).
• Cost-confirmation dialog pops up before the request so you can back out.

See AutoPrompt for the full flow and pre-flight gates (vision-model check, missing-figures friendly message, etc.).

Project persistence

The currently-loaded folder path is saved into the .svproj file, so reopening a project automatically re-loads its image context. If the folder has moved or been deleted, you get a warning in the log and the project opens with no images loaded — re-pick the folder via Load Folder to restore.

See also

• Prompt Manager — where Section 4 (Image Context) lives
• AutoPrompt — opt-in to ship figures with the meta-prompt
• AI Translation Overview — how images flow into per-segment translations