Canva Brush Edits in Place Without Showing What It Left Alone

Every one-change request carries a second request that nobody types: leave everything else alone. That second request usually carries the consequences. The design you're editing was already approved, and the parts the edit doesn't touch are supposed to carry that approval forward. If something else moved, the new version ships under a sign-off nobody gave it.

For most of the generative era, tools didn't try to honor the second request. They regenerated the whole output, and you rechecked the whole design. This week two products moved toward editing in place. A third, launched recently, shows how far "editable" output can sit from in-place revision:

• Canva shipped Brush on October 29. It is an editing model Canva built by taking an open-weight base model and training it further on design work. Brush edits by writing code against a design's elements, running it, inspecting the rendered result, and leaving everything editable.
• Google now lets Gemini users highlight text inside generated files (docs, slides, spreadsheets) and ask for a targeted refinement.
• Ideogram's Vector model outputs SVG files with real paths, marketed as editable.

My read is that revision has moved into the document and verification has stayed behind. None of these products gives the person who approved a visual design a way to confirm what the edit didn't change. That confirmation has two levels. The first is which elements moved. Brush's architecture makes that report cheap, and Canva hasn't built it. The second is whether the approved quality survived, which no list of changed elements can answer. Brush is the case to study on both counts. Canva's decision to train its own editor is also why the gap matters commercially: it signals that Canva means to own the edit while ChatGPT, Claude, and Gemini take over first drafts.

Where each product loses track of the original

Read the last column. Each of the three comparison products loses the original at a different point, while Brush's problem starts only after the edit is done.

Magic Layers breaks on the way in. Canva launched it in March to split flat images into editable layers. Its help page warns that complex and photorealistic images may not separate cleanly. User reports are anecdotal, but they line up with the warning. One Reddit user saw fonts and styles change during separation. Another got "metallic-looking blobs" from a product screenshot. Practitioner Hong Zhao got editable text out of a ChatGPT poster but cautioned that complex effects may not reproduce perfectly. None of that adds up to a failure rate. It does establish a dependency: whatever Brush preserves, it preserves relative to the rebuilt file. If the rebuild swapped your typeface, Brush will faithfully keep the wrong one.

Gemini breaks on the way out. Highlighting tells the model where to act. Google doesn't document whether the refinement stays inside the selection or rewrites the file. The larger issue is that the file you edit in chat is not the file you ship. Exporting to Docs creates a new document, and nothing says what survives the trip: fonts, pagination, formulas, speaker notes, comments. The edit is in place only inside the conversation; at the handoff it becomes a copy.

Ideogram never enters the revision loop. Its Generate SVG endpoint produces a fresh generation every time. The Vectorizer accepts raster images, not SVG. Masonry, which resells Ideogram access, inspected six outputs and found clean paths (7 to 2,001 per file), with lettering converted to outlines rather than live text. A one-word copy change in a generated logo therefore means redrawing by hand or regenerating. The output format improved, but revising it still happens outside Ideogram.

Brush knows what it touched

Brush works differently from all three. It writes a program, runs it against Canva's structured elements, then looks at the render.

Moderate confidence: if every edit is code executed against named elements, the system already knows which elements and which properties changed. A precise change list falls out of the architecture almost for free. Word can show tracked changes because computing the difference between two texts is easy. Brush is the first prosumer design editor in this group whose edits are structured enough to support the visual equivalent.

The interface doesn't surface any of it:

• Edits commit without approval. The Ask Canva help page says edits apply automatically. There's no propose-then-accept step before the change commits.
• Selection is a target, not a boundary. The same page says to select an element and prompt. It doesn't promise that nothing outside the selection will change.
• History doesn't identify the edit. Version history shows entries labeled "Canva AI," sometimes merged with your own changes. Nothing guarantees one saved version per instruction.
• Recovery is by eye. You can preview and restore versions. No highlight or list of changed elements is documented.
• The approver often can't look. Only owners and editors can open version history. In many teams the person who signs off has view or comment access.

Other tools do show what AI changed, at varying grain. Copilot in Word does word-level tracked changes, initially for Windows desktop Insiders. Figma's agent has Undo, and Figma's guidance for comparing before and after is to duplicate the design and undo the original, which is a workaround. Photoshop on iPad comes closest for visual work. It puts each generation on a separate non-destructive layer, so the edit sits somewhere you can hide, inspect, or delete. That boundary works because Generative Fill adds pixels on top of the image. Brush rewrites the properties of elements already in the design, so its change has no layer of its own to live on.

The real gap in visual tools, then, is granularity. Version history gets you the old design back without telling you what the edit did. In the Figma piece I called this recorded versus reviewed, and Brush lands on the recorded side. Canva's history records that AI acted. Reviewing what it did falls to whoever has editor access and a sharp eye.

What the launch numbers measure

Canva says Brush is "up to 69% faster for general design editing tasks" with "46% lower cost per edit task" than its previous design agent. Both figures are internal. The task set isn't published, "edit task" isn't defined, and the launch chart has no axis values. The cost figure is Canva's own cost per edit task, as Canva defines it. It says nothing about your subscription.

Neither number includes the minutes a human spends confirming that the rest of the design held still. In the persistent-agents piece I proposed counting supervisory minutes per safely completed goal. For creative revision, count reviewer minutes per accepted edit. An edit that runs twice as fast but still demands a full visual recheck saves the reviewer nothing.

There is not yet an independent hands-on test of Brush. Fast Company's coverage leans on Canva's demos and notes that one favorable social post was a paid partnership. Treat every claim about what Brush preserves as advertised behavior until someone measures it.

Approval lives in the space between layers

A Brush change list would answer the object-level question. The approval question sits one level up. Locks, selections, and change lists all operate on elements, while approval usually attaches to relationships between them: the vertical rhythm of a layout, the way type steps down through the hierarchy, the breathing room around a product shot, the color temperature held across a twelve-slide deck. No layer stores any of that, and a single-element edit is the edit most likely to break it. Lengthen a headline and the rhythm shifts while every other element stays technically untouched. An element-level report would read "headline changed, nothing else," which would be accurate even though the approval is broken.

Canva's protections work at the element level and are effectively all-or-nothing. Its Brand Template locks page says partial locks are treated as full locks during AI edits, which can block changes the template owner meant to allow. Brand Controls let admins restrict fonts and colors and require approval before publishing. Nothing documents whether Brush respects those restrictions, and the launch post lists Brand Kit support as future work. Whichever way that turns out, the structural limit holds: you can't lock a relationship without locking the elements that create it, and once those are locked the edit can't happen.

Adobe has gone furthest, and its approach marks the limit. Firefly Style IDs capture hierarchy, alignment, grouping, margins, and spatial relationships from approved Illustrator files. To prepare a file, you strip effects, distorted text, and drop shadows, so what gets captured is a simplified version of the approved quality. Brand Intelligence Validate checks layouts, typography, color, and imagery mood against configured guidelines and ties each flag to the rule it breaks. Adobe calls the web app a "functional reference implementation," configured per organization by its deployment team. Read it as a guided enterprise deployment, not a self-serve check any designer can switch on.

That leaves the industry with two instruments for distributed qualities: a blunt lock before the edit, and a configured brand checker after it. Neither lets a designer mark what was approved on this specific design and then get a check, after this specific edit, that it survived. Brush's loop already inspects the rendered page, which is where spacing and rhythm actually exist. In principle it could compare before and after renders across the whole layout, not just the element it changed. Canva doesn't say whether the loop checks anything beyond the requested change (speculative).

Why Canva trained its own editor

Canva gives two reasons for post-training its own model: it works more efficiently on design tasks, and it's cheaper to run at Canva's volume. Canva still mixes partner models with its own, but the in-house layer is growing. Its research group, formed in 2025 from the Kaleido and Leonardo teams, already counts Leonardo's Lucid Origin among its in-house image models, and on October 22 TechCrunch reported the release of another internally built image model.

Put that next to distribution. Canva has been inside Claude since July 2025 and ChatGPT since October 2025, and began rolling out in Gemini in May 2026. Canva's Gemini page documents a handoff from Gemini output into Magic Layers.

Moderate confidence: the strategy is to concede first drafts to the assistants and own the edit. Assistants increasingly produce the first version, often as a flat image. Magic Layers converts it into Canva's format, and Brush edits it there. The step that touches Canva's proprietary format now runs on a model Canva controls. Before, it depended on partners who are Canva's distribution channels and, with Gemini's file editing this week, its competitors in revision. The cost rationale fits the same picture: you train your own editor when edits are the high-volume business you plan to keep.

That makes the missing change report a strategic gap as well as an interface one. Assistants already draft, and as of this week Gemini revises files too. What they can't easily copy is an edit that accounts for itself inside a structured design file: which elements moved, under which locks, against the version someone approved. Only the product that hosts the file has the element model and the history to produce that account. If owning the edit is the plan, the report is what would make the position defensible, and Canva hasn't shipped it.

Two things weaken the picture:

• Routing is undocumented. Nobody has said whether Canva edits requested from inside ChatGPT, Claude, or Gemini run through Brush. If they don't, Brush becomes a reason to open Canva directly, which still serves the strategy (speculative).
• One handoff path is fragile. The assistant-to-Brush path Canva documents, the Gemini handoff, runs through Magic Layers. That's the step behind the reports of swapped fonts and the cautions about complex effects. On that path, Canva's newest model sits downstream of its least reliable conversion.

Verdict

• Interaction model: incremental. Select, prompt, apply is the pattern Photoshop Generative Fill, Ask Canva, and Figma's agent already established.
• Architecture: structurally new for prosumer design (moderate confidence, based on Canva's description). A post-trained model editing a proprietary structured format through code it runs and inspects is a different mechanism from repainting pixels or regenerating files.
• Trust layer: unchanged. There's no gate before the edit, since changes apply without a propose-and-accept step. There's no account after it, since history records only that AI acted, and only editors can see that. Relationship-level qualities are still protected only by locks that also prevent the edit.

Of the tools in this group, Canva built the design editor whose edits are most clearly structured enough to report themselves, and it shipped that editor without the report. Even a full report would stop at which elements moved; confirming the approved design survived needs a render-level check, and none of these tools offers one for a specific approved design.

What to watch

• A Brush-specific history entry or changed-element highlight. The cheapest signal that Canva will use what its architecture already knows.
• How Brand Kit support arrives. Enforcing Brand Controls and restricting edits is a different capability from feeding the model brand context. Read the release notes for which one ships.
• An independent before-and-after test. One requested change, with every other element and spacing value measured.
• Routing from the assistants. Whether Canva requests inside ChatGPT, Claude, or Gemini reach Brush.
• Adobe Validate's status. If it becomes a self-serve check that runs against a specific approved design after each edit, it becomes the nearest answer to the hard case.

For any creative-AI product you evaluate from here, ask how a reviewer learns what an edit changed and what it was supposed to leave alone. Every product in this piece can already aim a change, so that is where they will differ.