Presentations: AI Features

Create a new presentation project and choose Creator Mode. The AI wizard opens a multi-step flow:

1. Topic: Describe what the presentation is about in natural language.
2. Slide count: Choose the number of slides (or let AI decide).
3. Visual style: Select a preset style or describe your own (e.g. "dark, corporate, blue accents").
4. Generate: The AI creates all slides with titles, body text, speaker notes, and suggested layouts.

Merge mode recomposes 2–5 existing decks (and/or uploaded PDFs & PowerPoints) into a single, brand-new presentation — never a naive concatenation. Theo synthesizes a fresh Hook → Context → Evidence → Insight → Close narrative across every source, dedupes overlapping slides, and writes the merged deck around your instruction.

1. Open Presentations → New and switch the segmented control to Merge.
2. Add 2–5 sources: pick OpenCharts presentation projects with Add a presentation or upload PDFs/PPTXs with Upload PDF or PowerPoint.
3. Write a short merge instruction (e.g. “Merge into a 12-slide investor pitch keeping Deck A’s visuals”) so Theo knows what to optimize for.
4. Pick a visual-style source. Click Continue — Theo jumps straight to the outline step.

Clicking the + Add slide button in the thumbnail panel opens a focused modal where you decide exactly what the next slide should be before it lands on the deck.

• Theo, suggest the next slide: a Sparkles button that reads your existing slides (titles, bullets, speaker notes, and the original deck prompt) and proposes a natural follow-up concept. The suggestion drops into the textarea so you can keep editing it before submitting.
• Concept presets: one-click chips (Section divider, Key statistic, Process step, Closing thought) that seed the textarea with a starter prompt.
• Reference image (Creator mode): drop, paste, or upload an image. Theo classifies it and lets you choose whether to use it as the slide image directly or as a style reference threaded through the anchored cascade. Logos and icons default to the style-reference path so the asset never overwrites the whole slide.
• Visual style override:a condensed style picker defaulting to the deck or active slide's style. Pick a different look for just this one slide without changing the deck-wide style.
• Add blank slide: the escape hatch in the modal footer skips the prompt entirely and inserts an empty slide the same way the old flow did.

Once you submit, the new slide is inserted after the current one and — in Creator mode — immediately enters the existing rebuild shimmer until the image is ready. Keyboard shortcuts: ⌘ ↵ to add, Esc to close.

Each slide has a prompt bar at the bottom. Type a command to modify just that slide:

• "Add a chart showing revenue growth"
• "Make this slide more visual"
• "Simplify the bullet points"
• "Change the layout to two-column"

The AI analyzes the current slide content and applies targeted changes.

The per-slide AI bar accepts up to 12 reference images per turn, via any of three input methods:

• Upload: click the image button to open a file picker. Select multiple files at once — they all attach as separate references.
• Paste:⌘V/Ctrl+Vwhile the bar is visible. Works whether your cursor is inside the text input or anywhere else on the page.
• Drag-and-drop: drop image files onto the bar. A pink dashed overlay confirms the drop target.

Each newly attached image is analysed in parallel by a lightweight vision pre-flight so the bar can show contextual presets (Swap subject, Match style, Set as logo, etc.) tuned to what you actually uploaded.

Multi-reference editing (Creator mode):with 2 or more references attached, the AI never falls back to the "replace the whole slide image with a single upload" shortcut. Instead, it sends every reference image as labeled bytes (A, B, C, …) to the image model alongside your instruction, so you can swap multiple subjects in one shot:

• "Swap the three team members with these three reference photos, left to right."
• "Replace the woman on the right and the man on the left with these."
• "Use these images as combined style references."

Subject-swap rule: even with a single image, phrasing like "swap the guy on the right with this" or"change the woman in the middle"automatically routes the upload through the surgical reference path instead of overwriting the whole slide image. This is the fix for the historical "my upload took over the entire slide" behaviour.

Visual Edit Mode (Creator slides) lets you drag a marquee directly on the slide and either describe a surgical AI edit OR attach an image to drop into the selection. Click the Visual Edit button on the slide action bar to enter the mode — the cursor turns into a crosshair and the slide picks up a pink outline.

• Draw a marquee: click and drag a rectangle over the part of the slide you want to edit. The Edit Selected Region bar appears next to your selection.
• Text-only edit:type an instruction ("Change this title to say Executive Strategy") and hit Apply. Theo edits only the named area of the slide and preserves everything else pixel-stable.
• Attach an image:click the paperclip / Image button, drop a file onto the bar, or paste from your clipboard. A 36×36 thumbnail appears with a remove (×) button.

With an attachment staged, Apply routes the edit through one of three safe paths depending on what you selected:

• Image element swap (no AI): if your marquee covers a single image or logo element on the slide, Theo replaces its source directly. Instant, no credits, pixel-perfect.
• Logo composite (no AI): when the prompt or the upload looks like a logo, Theo composites your file deterministically into the region. Your logo lands crisp — the image model never gets to redraw it.
• Full replace: when the marquee covers most of the slide, Theo letterboxes the upload to 16:9 and uses it as the new slide image.
• AI blend: for smaller regions and natural-image references, Theo runs an anchored image-model cascade with the upload as a hard anchor for the named area, while preserving the rest of the slide.

Every successful edit snapshots the new output as the slide's pristine source, so subsequent AI edits start from a clean chain (no compounding artefacts). The slide's cumulative edit history resets after each region edit — same contract as the text-only flow.

The slide AI bar keeps an in-memory chat thread per slide. Every turn — your prompt plus a short assistant summary — stays attached to the active slide so the AI inherits the context on the next turn. You never have to restate which subject "the guy on the right" referred to, which reference replaced who, or what visual treatment you applied two turns ago.

• The thread expands inline above the input — the most recent 3–4 turns are visible at a glance.
• Click the New threadbutton on the thread header to clear the slide's memory and start fresh.
• Each slide has its own thread — switching slides switches threads, so context from slide 3 doesn't contaminate slide 4.
• Threads live in-memory onlyand reset on page refresh. They're a working memory for your editing session, not a permanent transcript.

What the AI sees on every turn: beyond your prompt and any attached references, the bar ships a structured slide contextblock so the model knows the composition it's editing — not just the rendered image:

• The slide title, bullet points, and speaker notes.
• The bounding-box geometry of every element on the slide (title, body, charts, images, shapes) so the AI knows what's where.
• The slide's layout, visual style, and theme colours (primary, accent, background).
• Lightweight previews of the previous and next slides plus the overall deck purpose, so cross-slide edits stay coherent.
• A compressed slice of the last ~10 turns of the slide's chat thread, so callbacks like"keep the same treatment, just brighter" resolve correctly.

Image elements have their srcdata stripped before being sent — only geometry and role are forwarded, so the model gets the spatial picture without re-uploading every embedded image on each turn.

Open the AI sidebar from the toolbar for deck-wide operations:

• Regenerate All Slides: rewrite the entire deck from a new prompt.
• Add Slides: generate additional slides on a topic and insert them at the end.
• Restyle: describe a visual style and the AI adjusts colors, fonts, and layouts across the deck.
• Summarize: condense the deck into fewer slides.

Click the Translatebutton (the languages icon in the toolbar) to convert the entire presentation into another language. Pick a target language and Theo rewrites every slide's text while keeping the exact same design, layout, colors, fonts, images, and transitions. Because translating the whole deck is a big change, Theo asks you to confirm first— in Creator mode the confirmation reminds you that every slide image will be rebuilt and redesigned with the translated text (which uses AI credits).

Supported languages:

• English
• Spanish (Español)
• French (Français)
• Italian (Italiano)
• Russian (Русский)

Translation covers slide titles, body text, bullet lists, quotes, chart labels, speaker notes, and the presentation title. Things that should never change — images, colors, fonts, icons, code blocks, and slide positions — are left exactly as they were.

The whole deck is translated in place as a single step, so one⌘Z(undo) reverts it if you want the original language back. The translated language is remembered on the deck and shown with a checkmark in the picker.

Translate a single slide: the floating toolbar above every slide (Creator and Content mode) has a Translate Slidecontrol. Pick a language to localize just that one slide — handy when you want a single slide in a different language, or to fix one slide without re-running the whole deck. In Creator mode the slide's image is rebuilt so the baked-in text shows the new language. Single-slide translation doesn't change the deck's remembered language, so a deck can intentionally mix languages.

Pick a language when adding a slide: the Add slide dialog has a Slide languagepicker. It defaults to the deck's current language, so adding a slide to an already-translated deck stays in that language — or pick a different one for just this slide. New slides are authored in English by default; choosing Spanish, French, Italian, or Russian localizes the slide as it's created.

Every deck ships with a per-deck image engine setting, controlled from the toolbar pill. Two engines are available:

• Theo Imagine (default): the premium engine. Higher resolution, stronger photographic and typographic fidelity. Best for hero decks, marketing material, and anything where image quality matters.
• Theo Creative: the fast engine. Lower credit cost per image and quicker turnaround. Best for early drafts, brainstorming, or large decks where speed beats fidelity.

Reliability: if Theo Imagine ever has a transient outage, your deck doesn't break — generation automatically falls back to Theo Creative for that one image and the editor surfaces a one-time toast so you know the swap happened. Switch back to Theo Imagine in the toolbar pill once the issue clears, or stay on Theo Creative if you prefer the faster engine.

Reliable framing: Theo Imagine slides are now generated with an explicit safe zone so titles and bullets always sit inside the visible 16:9 area. Slide images are also normalized to true 16:9 server-side, so the deck you see in the editor matches the file in your export pixel-for-pixel — no more text getting clipped at the top or bottom edge.

The engine setting also flows through the AI Image Generator modal (when you insert an image directly into a slide), so ad-hoc generations match the rest of the deck.

When you ask the Global AI Sidebar to rebuild the deck with a specific visual style, the AI locks every slide to that style instead of letting individual slides drift. Phrases like these are recognised automatically:

• "Rebuild every slide with cinematic style"
• "Make everything cyberpunk"
• "Use a watercolour look across the deck"
• "Switch to anime art on every slide"
• "Give it a noir film feel"
• "Vintage 80s VHS aesthetic"

The AI then rebuilds each slide one-by-one, applying the locked style to both copy and imagery so the whole deck stays consistent. Common synonyms — noir, neon, VHS, manga, cel-shaded, watercolour, DSLR, line art, geometric shapes, etc. — also map to the right canonical style.

Style picker apply-to-all: if you change the visual style from the picker and apply it to every slide, the editor offers to also rebuild the imagery so it matches the new style. Confirm to kick off the rebuild, or skip to keep your current images.

Logos behave differently from every other reference image. Theo never feeds a logo through the image model — the AI would inevitably reinterpret it (slightly wrong proportions, fuzzy typography, off-brand palette). Instead, logos are composited as a deterministic raster overlay on every slide so the asset stays pixel-perfect from slide 1 to slide N.

Canonical flow (recommended): open the slide's Properties panel → Brandsection → Upload Logo. This is the byte-for-byte deterministic path — your uploaded image is stored and rendered exactly as you uploaded it. No automatic background removal, no AI synthesis, no edits to the bytes. You can then pick a placement (top-left, top-right, bottom-left, bottom-right) and optionally toggle a rendering style: raw(default — paint the logo verbatim) or chip(wrap the logo in a translucent theme-aware glass backdrop, helpful when an opaque white PNG would otherwise read like a sticker on a dark slide). Both options are opt-in — we never flip Chip on for you.

Tip:if your logo has a tinted background you want to drop, export it as a transparent PNG from your design tool before uploading. Theo will not auto-key the background — it preserves whatever you ship.

Global AI “use this everywhere” also routes through the safe path.When you attach a logo image to the Global AI composer and ask Theo to keep it consistent, the vision classifier flags it as a logo, the server pins it as a logo-mode brand anchor, and the editor automatically composites it via the same overlay used by the Brand Section upload — never through the image model. Trigger phrases Theo recognises:

• “Use my logo on every slide”
• “Keep my brand mark consistent”
• “Add the company logo”
• “This is our logo — pin it”
• “Apply the corporate badge”

Per-slide references: logos are refused. The Slide Image Studio modal detects logos in the References tab and refuses to anchor them onto a single slide — instead, an inline banner offers a one-click Set as deck logoCTA that routes the asset through the same deterministic overlay. If you still hit Save & Rebuild while a logo reference is staged, the rebuild payload strips it out so the logo bytes never reach the image model.

Pinning a logo as a brand anchor (via Global AI) is treated identically to a Brand Section upload — the asset becomes theme.logoUrl and every slide's role: "logo" element updates in place. The rendering style (raw vs chip) honours whatever you previously picked in the BrandSection toggle, defaulting to raw. Reload-safe: the apply only fires when the pinned URL differs from the deck's current logo, so re-opening a deck with the same logo is a no-op.

What about Theo Imagine / Theo Creative drift?Image models can still hallucinate fake logos, wordmarks, signatures, monograms, or brand badges in the corners of a slide if asked broadly. Theo's safe-zone directive explicitly forbids the image model from rendering ANY of those, so the reserved corner stays empty for the deterministic overlay to composite on top. If you ever see a duplicated logo on a slide, click Rebuild— the second pass will inherit the strengthened directive and clear the imposter.

Auto-route when you skip the trigger phrases. Natural-language prompts that don't match a logo phrase (e.g. “anywhere you see open charts, use this”) still hit the deterministic path. The Global AI server runs a vision classifier on every attached image and — when it flags the upload as a logo and you didn't explicitly ask to pin it as a brand anchor — emits an appliedAsDeckLogo signal. The editor composites the asset as a corner overlay on every slide and scrubs any “render the {brand}logo prominently” sentences out of the per-slide image prompts so the rebuild loop cannot paint a fake brand mark inside the imagery.

Targeted logo placement: only on slides that match. If you want the logo on a SUBSET of slides instead of every slide, just say so. Theo recognises targeting language and switches from the deck-wide overlay to a per-slide placement. Examples:

• “Use this logo, but only on slides that mention OpenCharts.”
• “Add the logo anywhere the slide says ‘Theo’.’”
• “Pin the company badge for those slides only — not the whole deck.”
• “Don't use it for all slides — just the ones about pricing.”

When Theo detects targeting language plus a logo upload, the deck-wide overlay path is suppressed and the planner reads every slide's title and body to find the matching slides. Each match gets a per-slide role: "logo" element using the same deterministic raster overlay — never the image model. Unrelated slides stay untouched, and theme.logoUrl is not modified, so the deck-wide overlay does not fire for everything else. You can drag, resize, or remove the per-slide logo on each slide directly.

When no slides match.If the planner can't find any slide that matches your criteria from the deck content, Theo surfaces a warning toast (“No slides matched your targeting — nothing changed”) and leaves every slide untouched. Rephrase the targeting (e.g. name a more specific keyword that you know appears on at least one slide) or open a single slide's AI bar to add the logo there directly.

Refresh existing imagery (one click). When the auto-route fires in a Creator-mode deck that already has rendered slides, Theo offers a confirmation dialog: “Some slides may still show an older AI-painted brand mark inside the image. Rebuild them now?” Confirm to refresh every Creator slide so the underlying scene matches the new overlay; skip to keep your existing imagery and rely on the corner overlay alone.

When you attach an image to Global AI and ask Theo to keep it consistent across the deck, Theo pins the image as a Brand Anchor. From that turn on, every Creator-mode slide regenerates with the actual image bytes feeding the image model — instead of being reimagined from a text description that drifts further each slide.

Three application modes:

• Subject anchor — the image IS the literal subject (your mascot, character, product hero shot). Every rebuilt slide preserves identity, silhouette, proportions, colors, and core design. Only pose, framing, lighting mood, and surrounding scene adapt to the slide topic. Theo will not silently add wings, capes, accessories, or restyle the subject unless you explicitly ask.
• Style reference — the image defines the visual style (palette, lighting, rendering technique). Theo renders the subject from your slide directive while matching the reference’s look.
• Logo — deck-wide logo replacement. Composites the asset as a corner mark across every slide via the existing logo pipeline (the image bytes are not fed to the image model).

Trigger phrases Theo recognises: “use this image”, “use this mascot”, “keep consistent”, “pin as anchor”, “same character”, “don’t change”, and similar variations. Once pinned, the anchor persists on the deck so single-slide rebuilds and “Rebuild all” also use it — you do not have to reattach the image every turn.

Drift-reduction guardrails (subject mode): when a subject anchor is active, Theo runs a multi-engine cascade tuned for identity preservation, in this order:

1. A purpose-built identity-preserving image edit pass that takes your reference as the literal source frame and only adapts the scene around it.
2. A high-fidelity multi-image render with the reference (and an isolated background-removed cutout of your subject) sent directly to the image model alongside the slide directive.
3. A faster multi-image render as the last resort, with the same reference + cutout payload.

Sampling is locked low so identity-defining features (silhouette, palette, accessories) stay stable across rebuilds. The accompanying slide prompt is automatically trimmed to a short composition-only sentence so the verbose theme palette and safe-zone rules can't overwhelm the reference image. Style anchors skip step 1 because they’re matching aesthetics, not a literal subject.

Drift verifier + bounded retry: after each anchored slide renders, Theo compares it back against the pinned reference and scores how faithful the result is. Clean matches ship silently. Visible drift (added accessories, wrong palette, restyled subject) triggers ONE hardened retry that names the drift back to the model so it can correct on the next pass. If the retry still drifts, the slide ships marked anchor weakso you know which slides may want manual attention.

Fallback behaviour: if the reference bytes can’t be loaded or every anchored engine fails, the slide quietly falls back to the standard text-only cascade and the editor surfaces an anchor dropped warning so you can rebuild manually. Logos always use the deterministic logo pipeline, never the anchored image-model path.

The Slide Image Studio (toolbar → Edit Prompt/ Style) is the canonical place to fine-tune one slide's image prompt, attach 1–6 reference images, pick a visual style, and choose which engine renders the rebuild. Unlike the deck-wide Brand Anchor, Studio references apply to just the current slide and the FULL slide text (title + bullets + theme typography) is rendered back into the image.

Three reference modes— each maps to a different image-model contract:

• Subject— preserve the literal subject pixel-for-pixel. Best for mascots, products, and repeating characters where identity must stay locked across rebuilds.
• Concept— NEW. Use the reference as inspiration only. The slide preserves the archetype, role, vibe, and count of subjects (e.g. “three people working at a table”, “a confident speaker on stage”) but Theo is free to reinterpret faces, outfits, palette, and rendering medium in the slide's own visual style. Best when you upload photos of real people for inspiration rather than literal replication.
• Style— match palette, lighting, and rendering technique only. The slide's subject is free to be whatever the prompt asks for. Best for moodboards and aesthetic references.

When you drop a reference, Theo classifies it (People / Mascot / Product / Scene / Logo / Image) and shows a Detectedbadge on the thumbnail. The mode toggle defaults to the best match — people → Concept, mascots and products → Subject, scenes → Style — and you can override at any time.

Per-rebuild engine picker— Auto, Theo Imagine, or Theo Creative:

• Auto— inherit the deck-wide image engine. Pick this when you don't need a one-off override.
• Theo Imagine— the premium engine. Strongest at preserving real people and rendering slide text inside the image. Use it when the reference is a photograph of a person you want to keep recognisable.
• Theo Creative— the fast Gemini-based engine. Best for mascots, illustrations, and stylistic edits where you want speed + creative reinterpretation over photographic fidelity.

Both modes and engine flow through Save & Rebuild for the current slide only — they don't alter the deck-wide Brand Anchor or imageEnginesettings. Theo respects the slide text on every rebuild so titles, bullets, and theme typography survive instead of getting stripped out for an image-only render.

The Slide Image Studio's Referencessection now opens to a three-tab strip so you can mix sources without jumping between modals. Every tab feeds the same staged-reference grid underneath, so the first staged item is always the primary anchor for the rebuild — mode toggles, engine picker, and Save & Rebuild work exactly the same regardless of where the reference came from.

• Upload — the existing drop-zone / file picker / paste handler. Up to six references can be staged in parallel; the classifier tags each one (People / Mascot / Product / Scene / Logo) so the default mode lands somewhere sensible.
• From this deck— NEW. Shows a tile for every slide in the deck that already has imagery (excluding the slide you're editing). Click a tile to stage it as a style reference — no upload step, no duplicate file in storage. Re-click to remove it again. Use this when you want a new slide to match the palette, lighting, and aesthetic of a slide you've already built.
• Mascot — NEW. A first-class place to pin a deck-wide brand subject (your mascot, recurring character, hero product). Pinning a mascot stores it on the deck so every NEW slide added afterward inherits it automatically, and a one-click Use on this slide stages it for the current rebuild with the subject-preservation cascade.

Non-destructive contract for the Mascot tab. This is the most important guarantee — pinning, replacing, or removing a mascot will never silently rewrite slides you've already rendered:

• Pinning affects only (a) the rebuild you fire next with Use on this slide on, (b) every NEW slide added afterward, and (c) batch rebuilds you explicitly opt into via Apply to all slides… (confirmation dialog).
• Replacing the mascot swaps the pinned asset in place. Already-rendered slides keep their current images until you rebuild them.
• Removing the mascot clears the pin. Nothing is rewritten or deleted. Future slides simply stop inheriting it.

When the Mascot is staged for a rebuild, Theo runs the identity-preserving anchored cascade (subject mode) so the mascot stays the literal subject of the slide. Title typography is rendered prominently in-image; bullets and long body copy are intentionally softenedon these rebuilds. Image models reliably garble dense paragraphs when they are also locked to a reference subject (you get duplicated or misspelled words), so Theo will summarise body text to a short supporting tagline or omit it entirely — the mascot + title combination wins every time. If you need the full body text rendered into the slide, run the rebuild without the Mascot staged and re-add the mascot afterward.

Any extra references you stage from Upload or From this deck on the same rebuild still flow through as supplementary cues (composition / palette / mood) without overriding the mascot.

In the slide background picker, click the AI sparkle button and describe the background you want. Examples:

• "Abstract blue gradient with subtle geometric shapes"
• "A blurred photograph of a city skyline at sunset"
• "Minimalist pattern with thin gold lines on dark navy"

The AI returns an image that is automatically set as the slide background. You can regenerate or edit afterward.