Interstitial AI Build log

June 2026 update

Tue, 30 Jun 2026 00:00:00 GMT

In June our first pilot studios converted into paying B2B accounts. We've had individual subscribers since January, but studios paying as a team is new, and it set the month's priorities. Most of the work went into the billing a studio needs. Alongside that, we brought regional editing back at full resolution, rewrote how groups behave, gave every image layer proper color adjustments, and built an asset library for keeping and reusing your own references.

June in numbers

Our first paying studios. Pilot studios converted to paid B2B accounts, and more firms started pilots.
11 new test suites for billing. Invoicing, seats, and the shared balance.
9 element types, one set of group rules. Every kind of element now behaves the same inside a group.

Shipped

Self-serve studio billing. A studio can now subscribe on its own, with no pilot and no sales call. From settings it can add and remove seats as the team changes, cancel, and come back later. Each member gets their own monthly credits, and the whole studio shares one balance on top for whoever needs it. Payment is by card, SEPA direct debit, or bank transfer. With real studio money involved, we made the billing careful and predictable. Nothing runs without being paid for, and a studio can cap how much each member spends.

Regional editing, rebuilt. We said in May this was coming back. Rather than patch the old version, we rebuilt it. You mask the part of an image you want to change, and it now works on the full-size image, up to 4K, so the edit is sharp and lands exactly where you masked it. A live outline shows the region you're about to change. You can guide the edit with reference images and generate in batches. Layer masks were part of the same rebuild. A single mask now holds both hard and soft edges, you can copy and paste masks between layers, and you can pan the canvas while you work.

Groups, rewritten. Groups already existed, but they behaved differently for every kind of element, and testing with users made that plain. We rewrote them against a single set of rules. Images, drawings, notes, text, cutouts, and every kind of widget now group, move, copy, and delete the same way, nested up to five levels deep. A group can carry a color and be hidden or locked. Drag a loose element onto one and it joins. Delete a group and its contents go with it. Empty a group and it dissolves on its own.

Per-layer color. Every image layer can now carry Curves, Levels, and a stack of Photoshop-style adjustments, and they update live as you drag. Adjustments survive a reload, and when you generate, the AI works from the color-adjusted image you see, not the original.

Generation. Nano Banana Pro and Nano Banana 2 can now pull in real references from Google while they work. It's off by default, and adds a small extra cost when you switch it on. Nano Banana 2 also gained a thinking level for harder prompts. We fixed two problems with 4K renders: they could be slow, and they could come out at a lower resolution than you asked for.

Workspaces. Every project tab is now just a canvas you drop layers onto. The old fixed "scene image" is gone. Uploads and CAD or Revit extractions all arrive as ordinary layers, and an extraction lets you choose which workspace it lands in.

Asset library. A shared library for your images and PDFs, personal and studio-wide. Save into it from the canvas, your clipboard, or straight from Are.na. Organize with nested folders, tags, ratings, and notes. Filter by color, aspect ratio, or type, search it, and drag anything back onto the canvas. A studio can publish references to a shared library with per-member permissions, so a firm keeps one set of references everyone works from. Save the same image twice and we keep only one copy, so duplicates cost nothing.

On the road

Auer Weber. We ran a workshop at the firm, and the most useful part was them showing us how they use Interstitial on their own projects. We were surprised by how many different uses they had found for it.

Tongji University, Shanghai. Architecture students in TUM's Digital Futures Workshop at Tongji are designing and rendering with Interstitial. It's our first classroom outside Europe, and it's still running.

KIT and TUM. We presented to the architecture master's students at KIT, and our teaching at TUM continues.

Learned

Studios raise the stakes on billing. We've charged individuals since January, but a studio is a different kind of customer: bigger invoices, several seats, a signed contract behind them. That gives billing far more surface to get right, and at this scale a billing error would cost a firm real money and trust. It cannot happen, so most of June's billing work went into making sure it never does.

Some things are worth rebuilding, not patching. The fast way to bring regional editing back was to patch the old pipeline and keep its wrong-resolution masks and smeared edges. Rebuilding took longer, and it is the only reason the feature is worth using.

Plugins. Rhino, Archicad, and SketchUp, plus a large update to the Revit plugin, so models come straight from CAD into Interstitial.
New models. The latest image, video, and segmentation models, added as they land.
Sharing. Read, collaborate, or copy access, shared by a link, outside the firm or inside your own studio.
Hardening. Bug fixes and performance work across the board.

Interstitial is an AI-driven architectural visualization platform built in Munich. We send build updates monthly. Contact

May 2026 update

Sun, 31 May 2026 00:00:00 GMT

In May we shipped the editor features users had been asking for: undo, batch generation, and faster loading. We turned on company billing, so studios can pay as a team and move from pilot to paying account. We rebuilt Design Mode around a chat agent. We started pilots with more architecture and interior design firms. And the firms that spent their pilots comparing us with our competitors are choosing us.

May in numbers

10 phases shipped. Planned, executed, tested, deployed.
119 new tests. 96 frontend, 18 backend, 5 end-to-end.
Worst-case canvas freeze cut from 16.7 seconds to 0.27. Measured on a 4K upscale.
1,300 lines of legacy code removed. The old Fabric.js drawing engine, dropped.

Shipped

Undo and redo. Ctrl+Z and Cmd+Z now work across the canvas: layer moves and deletes, vector placement, handle drags, rotate, scale, and toolbar edits. We rebuilt the edit path so every change records its own inverse and adds one step to the history. Drag five layers at once, undo, and all five go back together. The same work let us remove the old Fabric.js drawing engine. It was our most-requested feature.

Company billing. Paid organizations now share one credit pool with proper multi-seat invoicing. Each seat gets its own tier, and we handle the full Stripe lifecycle: invoices, subscriptions, tax, and a dunning flow for failed payments. German VAT is built in, including Kleinunternehmer status and PAngV price disclosure. We also simplified the pricing and added a magic-link cancellation page, so leaving never needs a support ticket.

Batch generation. Pick a count, press Generate, and we run the requests together, drop the results into a stack, and expand it into a grid once there are two or more. New images now land in a placement grid next to their source instead of on top of the original, which keeps the workflow moving. The batch is saved as it runs, so reloading partway through never loses a result.

Performance and reliability. We rebuilt canvas image loading around one queue with a concurrency cap: thumbnails load first, full resolution fills in behind. We moved image encoding into a worker off the main thread, which cut the worst-case freeze on a 4K upscale from 16.7 seconds to 0.27. We also built a defense against a database crash that hit when the browser's storage filled mid-session: it recovers instead of failing, and refuses to reload when that would cost unsaved work. 62 tests cover it.

Design Mode chat. Design Mode is our flow for rendering CAD drawings and photos. In April it could classify your image and walk you through a fixed menu. In May we gave it a chat agent. It recognizes five kinds of input, tells twelve kinds of architectural drawing apart (elevations, sections, axonometrics, and more), and renders each in its own style while keeping technical drawings on their true view. It keeps a separate chat thread for each image and remembers where you left off. That is about sixty input situations, up from roughly nine.

Model reliability. We added no new models this month. We made the dozen image and video models we already run faster and more reliable, and moved our whole Gemini stack onto Google's stable endpoints before the preview ones retire on July 17, with no interruption. We also fixed an empty-response bug on Gemini, fixed GPT Image 2 output that was finishing but never reaching the canvas, and removed a "Veo Fast" label that pointed at the standard model.

On the road

RenderTube workshop and Testat I at TUM. We ran another workshop in TUM's RenderTube course, where architecture students use Interstitial to learn visualization, and we came back as guest critics for their Testat I reviews. Teaching architecture students is also some of the fastest feedback we get.

Google hackathon. Google invited us to a hackathon for its business partners and a few startups, with their engineers helping us work through our own problems. We were one of the youngest companies there. We used the time to push on our agent system, which we want to grow from a design helper into an assistant that can understand and operate the whole app. We also found inefficiencies in our cloud and storage setup that we plan to fix soon.

Learned

B2B work is unglamorous and necessary. Bilingual GDPR documents, a German imprint, self-service cancellation, two-click confirmations on destructive admin actions. None of it is exciting, and all of it is what a firm checks before it signs. May had a lot of it.

Some users find us through Claude. A few new users told us they found us by asking Claude for a tool to visualize architecture, and it recommended ours. We did not pay for that and cannot, but it is becoming a real way people find the product.

A real agent. Growing the design helper into an assistant that can understand and operate the whole app. This is what we worked on at the Google hackathon.
Plugins. Our Revit plugin gets Revit 2026 support, on top of the 2025 version it runs in today. We are starting work on Archicad and Rhino, with more to follow.
Pipeline canvas. Connecting generations into reusable pipelines, right on the canvas.
Bigger sharing. Read, write, or copy access, shared by a link or inside your own studio.
Colormaps and alpha maps. Colormaps you can upload and download for automatic segmentation, and alpha maps for masking.
Regional editing. Coming back, with pixel-perfect control over what gets edited and what does not.
Co-editing on one canvas for up to twelve. We have already built and shipped the backbone that lets multiple people work on the same canvas at once. Now we are turning it into full co-editing, for up to twelve people at the same time.
Leaner cloud and storage. Acting on the inefficiencies the Google hackathon surfaced.

Interstitial is an AI-driven architectural visualization platform built in Munich. We send build updates monthly. Contact

April 2026 update

Thu, 30 Apr 2026 00:00:00 GMT

April was our first month live at established architecture firms. They started evaluating us alongside our direct competitors. We blocked a three-day attack on free signups and built fraud detection in response. We finished the multi-tenant infrastructure we need to sell to studios. We launched publicly in Munich and started a guest-lecture series at TUM and KIT.

Watch the TUM lecture: https://youtu.be/YopWSnXx75w?si=9DsH3Yc2ScO3czPP

April in numbers

10 phases shipped. Planned, executed, tested, deployed.
90+ malicious accounts blocked. 70+ manually during a three-day attack, 20+ automatically since the fraud system shipped. Two distinct attackers.
8 generative AI models in production. Seedream, Flux 2, NanoBanana 2, NanoBanana Pro, GPT Image 2, Kling 3.0, Veo, Seedance 2.0.
3 university partnerships. TUM, KIT, and Frankfurt University of Applied Sciences.

Shipped

B2B pilot infrastructure. Studios can now run multi-seat pilots with role-based access, shared workspaces, project-level sharing, and per-organization usage tracking. We rebuilt the authorization model around organizations: shared-viewer auth on private storage, org-aware tier gating across the generation pipeline, and atomic flows for invites, pilot conversion, and end-of-pilot cleanup. Future team features build on this.

Account security and fraud detection. Built in response to the signup attack (more below). It includes phone verification, Cloudflare Turnstile, an automated block-and-appeal flow, a Gemini-based fraud-analysis agent that reviews signup signals before approving accounts, and a few additional safeguards we'll keep to ourselves. We also built an internal admin tool to triage flagged accounts and process appeals.

Smart Select. Browser-based object selection on the canvas, using Meta's Segment Anything Model 2 (SAM2) via onnxruntime-web. Click on a chair, a tree, a window, or a facade and it gets selected. No manual masking. Tiered model selection keeps it fast on lower-end hardware.

Generation model refresh. Added GPT Image 2 (live the day OpenAI released it), Kling 3.0, and Seedance 2.0. Veo got a Blueprint-aware switcher and moved to its stable GA model ID. Video generation now uses per-second credit pricing, so users pay for what they actually generate.

Canvas: stacks, soft masking, editor redesign. Image stacks now have an expanded masonry view with drag-in/drag-out reorganization, deep-zoom direct manipulation, and a Photoshop-style soft brush for layer masks. The editor icon system was redesigned. New selection tools, canvas panning.

Onboarding, tutorials, and sharing. Five in-product tutorials for the most-used workflows (edit-image annotations, reference objects, design mode for CAD, design mode for photo, facade rendering). Public copy-link project sharing. A rebuilt onboarding flow built around workspaces instead of the older scene-upload model.

Partner-university access. Students and faculty at TUM, KIT, and Frankfurt University of Applied Sciences now get enhanced Blueprint access. Phone verification is waived.

On the road

Munich launch event at ANÍCE.STUDIO. Our first public showing of the platform to architects, designers, and the local startup community.

TUM and KIT guest lectures. We started giving lectures on AI in architectural visualization at two of Germany's leading technical universities. The first TUM session is linked above. KIT is in person.

Learned

Established firms see us on the same level as our direct competitors. They're running early evaluations against them. No final calls yet. The reason we're holding even is workflow fit. Canvas-and-stacks and multi-image reference handling map to how architects actually work.

Bad actors find you fast. Within a week of advertising our free tier, an attacker in Egypt started mass-registering accounts to drain free credits. We blocked over 70 of their accounts manually over three days while we built defenses. They adapted to each new check we added. By the end we had a layered system: phone verification, Turnstile, an automated block-and-appeal flow, a Gemini agent that reviews signup signals before approving accounts, and a few additional safeguards we'll keep to ourselves. Since deployment, 20+ more accounts blocked from two distinct attackers. Zero successful malicious registrations.

Scaling reveals what was fragile. Production use surfaced reliability issues no test environment catches: generation timeouts, stale recovery states, signup memory pressure, payment-renewal edge cases under Stripe's new API. We fixed each one as it came up. The platform is more reliable than it was a month ago.

Company billing. Multi-seat invoicing for paid organizations, on top of the pilot infrastructure we shipped this month.
Undo (Ctrl+Z / Cmd+Z). Long-requested. Finally landing.
Batch generation in expanded stacks. Run multiple generations in one click.
Better generation placement. Smoother workflow, fewer interruptions, more productivity.
Performance, loading, and storage optimization. Faster cold starts, lighter bundles, smarter asset handling.
Collaboration and sharing. Real-time co-editing on the canvas.
Plugins for Rhino, Archicad, and SketchUp. Models flow directly from CAD into Interstitial. No export-import step.

Interstitial is an AI-driven architectural visualization platform built in Munich. We send build updates monthly. Contact

Interstitial AI Build log

June 2026 update

June in numbers

Shipped

On the road

Learned

Next

May 2026 update

May in numbers

Shipped

On the road

Learned

Next

April 2026 update

April in numbers

Shipped

On the road

Learned

Next